Some Further Evidence on Exchange-Rate

Volatility and Exports

George Hondroyiannis, P.A.V.B. Swamy, George Tavlas, and Michael Ulan

Bank of Greece and Harokopio University; U.S. Bureau of Labor Statistics;

Bank of Greece; U.S. Department of State (retired)

Abstract: The relationship between exchange-rate volatility and aggregate export

volumes is examined using a model that includes real export earnings of oil-expor-
ting economies as a determinant of export volumes of a sample of 12 industrial
countries. Four fixed-coefficient panel-data estimation techniques, including
a generalized method of moments (GMM) and random coefficient (RC) estima-
tion, are employed on panel data covering the estimation period 1977:1–2003:4
using three measures of exchange-rate volatility. Our aim is to provide a theoret-
ically and empirically justifiable specification that can guide researchers. In con-
trast to recent studies employing panel data, we find little evidence that volatility
has a negative and significant impact on trade. We use second-generation RC esti-
mation, which corrects for biases arising from incorrect functional forms, omitted
variables, and measurement errors. Our results suggest that the finding of a sig-
nificant and negative impact of volatility is attributable to specification biases.
JEL no. C23, F3, F31
Keywords: Exchange-rate volatility; trade; random-coefficient estimation; general-
ized method of moments; panel data

1 Introduction

A large empirical literature has investigated the relationship between

exchange-rate volatility and trade flows. While the earlier literature (circa
1980 to the mid-1990s), employing mainly time-series data and ordi-
nary least squares (OLS) estimation, did not, by-and-large, find a negative
and significant effect of exchange-rate volatility on aggregate trade vol-
umes, recent studies, often using panel data and more-elaborate estimation

Remark: The views expressed are those of the authors and should not be interpreted as
those of their respective institutions. We are grateful to Stephen Hall for helpful com-
ments. An anonymous referee provided thoughtful comments and questions which helped
us improve the paper. Please address correspondence to George Tavlas, Economic Research
Department, Bank of Greece, El. Venizelos 21, 102 50 Athens, Greece; e-mail: gtavlas@
[Link]

© 2008 Kiel Institute DOI: 10.1007/s10290-008-0141-4

152 Review of World Economics 2008, Vol. 144 (1)

techniques (fixed effects, random effects, generalized method of moments

(GMM)) and specifications of volatility (e.g., generalized autoregressive
conditional heteroskedasticity (GARCH) models), have uncovered some—
though by no means overwhelming—evidence of a negative, significant
relationship.1 For example, in a study that surveyed recent work and pro-
vided new evidence, Clark et al. (2004: 3) found that “some recent studies,
as well as some of the evidence presented here, appear to suggest that the
data support a negative relationship”.2
This study aims to shed light on the differences in results obtained in
earlier studies, taken as a group, and more recent studies (taken as a group).
Our point of departure is the analytic framework proposed by Bailey, Tavlas,
and Ulan (1986) (henceforh BTU). Those authors investigated the rela-
tionship between exports of the seven largest industrial economies and
the short-term volatility of the exchange rates of the currencies of those
economies based on a specification that included real export earnings of
oil-exporting economies as a determinant of export volumes. The suppo-
sitions underlying this specification were as follows: (1) since the 1970s,
oil-exporting economies have provided important markets for exports of
industrial economies, and (2) given their levels of economic development,
the income elasticities of demand (with respect to their real export earn-
ings) of the oil exporters for industrial-country exports might well differ
from the income elasticities of demand for those goods in other industrial
countries. BTU found that oil-exporters’ elasticity of demand (with respect
to their real export earnings) for industrial-country exports was statistically
significant. In light of the large fluctuations in oil prices (and hence in
revenues from oil exports) in recent years, the inclusion of such a variable
appears to be especially appropriate in present circumstances. Using two
measures of unconditional exchange-rate volatility, involving the absolute
values of quarterly percentage changes in effective exchange rates of the
countries concerned, BTU found no evidence of a negative and significant

1 In some cases, recent authors found a significant and positive relationship between
exchange-rate volatility and trade volumes. Surveys of earlier literature include IMF (1984)
and Edison and Melvin (1990). The more recent literature is surveyed by McKenzie (1999),
Clark et al. (2004), and Coric and Pugh (2006). The latter study found that, of 49 studies
surveyed, 29 studies reported that exchange-rate volatility reduces trade.
2 The authors of that study were careful to point out, however, that such a negative rela-
tionship was not robust to all specifications they estimated. In his survey, McKenzie (1999)
similarly found that the authors of recent papers appeared to be having greater tendency
to obtain such a [i.e., negative] relationship.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 153

effect of volatility on trade for the seven industrial countries considered in

their study.
This paper deals with the following issues:
(i) To what extent were BTU’s results driven by those authors’ choice of
a particular set of seven countries and a particular estimation tech-
nique (i.e., OLS)? To address this issue, we follow the recent litera-
ture by applying constant-coefficient panel-data estimation methods
to the original group of seven industrial countries studied by BTU as
well as to an expanded group of 12 industrial countries. Specifically,
we use common-fixed-coefficient estimation, fixed-effects estimation,
random-effects estimation, and GMM estimation applied to a panel-
data set. The GMM approach, proposed by Hansen (1982), does not
require distributional assumptions, such as normality, can allow for
heteroskedasticity and serial correlation of unknown form in exactly
identified cases, in which the number of moment equations is the same
as parameters to be estimated (Greene 2003: 548), and purportedly
can correct for the effects of specification errors including omission of
relevant variables (Verbeek 2004: 148–153).3
(ii) Much of the recent literature has used conditional volatilities, such as
GARCH, to proxy exchange-rate volatility. To what extent have recent
findings of a significant impact of volatility on trade been influenced
by the particular measure of volatility used? To help answer this ques-
tion, we employ a GARCH measure of volatility in addition to the
unconditional measures used by BTU.
(iii) Were BTU’s findings driven by the particular model (involving export
earnings of oil exporters as an explanatory variable) used? To examine
this issue, we dropped the variable representing real export earnings of
oil exporters and re-estimated regressions under each of the estimation
techniques employed.
(iv) What are the effects of specification biases in a particular model on
the results obtained? To address this issue, we use random coeffi-
cient (RC) estimation. The RC approach deals with four major spe-
cification problems (discussed in Chang et al. 2000) that almost al-
ways arise in econometric estimation.4 RC estimation comes in two
varieties: first-generation and second-generation models (Swamy and

3 Having said that, the assumptions underlying the GMM approach are questioned below.
4 For discussions of RC estimation, see Swamy and Tinsley (1980), Swamy et al. (2003)
and Swamy and Tavlas (2001, 2005, 2006, 2007).
154 Review of World Economics 2008, Vol. 144 (1)

Tavlas 2001). First-generation models correct for misspecifications of

functional forms. Second-generation models take as their point of
departure the premise that, although one can never be sure that a “true”
model (in this case, the “true” model of the determinants of a country’s
exports) exists, RC estimation, by correcting for factors that cause spu-
rious relationships (e.g., the effects of omitted variables, incorrect func-
tional forms, and measurement errors), can find the most-reasonable
approximations to the “true” values of the identifiable coefficients of the
“true”, but unknown, model.5 In what follows, we apply both variants
of RC estimation. As we discuss below, use of second-generation RC
estimation provides a basis to search for a specification that performs
well in both explanation and prediction (Zellner 1988).
Briefly to anticipate, using the fixed-coefficient panel-data methods
commonly applied in the recent literature, we find no evidence of an effect
of exchange-rate volatility on exports of the countries considered, regard-
less of the measure of volatility used. However, using a specification that
drops the variable representing real export earnings of oil exporters, we find
some evidence of a significant and negative effect of volatility when using
GMM estimation and a GARCH measure of volatility, a result consistent
with that found in much of the recent literature. Which specification, then,
that with the variable representing export earnings of oil exporters or that
without that variable, is the appropriate one? To help discriminate between
the specifications, we use RC estimation. Second-generation RC estimation
is particularly suitable in this case because it directly confronts omitted-
variables bias; if export earnings of oil exporters are an appropriate variable
in the trade equations, its omission is dealt with under second-generation
RC estimation in that the “true” coefficients on the included explanatory
variables are captured. We find that the effect of volatility on export volumes
is not significant under RC estimation, suggesting that the finding of such
a significant effect elsewhere in the literature is attributable to specification
biases not captured in constant-coefficient panel estimation procedures,
such as GMM.
The remainder of this paper consists of six sections. Section 2 briefly
summarizes recent studies that use panel-data estimation. Section 3 is an

5 The coefficients of the “true” model are treated as the “true” coefficients which are free
from incorrect-functional-form, omitted-variable, and measurement-error biases. Only the
“true” coefficients on the explanatory variables included in a specified model are identifi-
able on the basis of the available data (Swamy and Tavlas 2006: 419, 2007).
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 155

overview of the theoretical relationship between exchange-rate volatility

and trade. Section 4 discusses the model and data. Section 5 provides a brief
description of the estimation techniques. Section 6 presents the empirical
results. Section 7 concludes.

2 Literature Review

Previous studies employing panel data have tended to find evidence of

a negative and statistically significant relationship between exchange-rate
volatility and trade. Wei (1999) estimated a panel of 63 countries over the
years 1975, 1980, 1985, and 1990; he examined a total of over 1,000 coun-
try pairs. Using switching regressions, the author found that, for country
pairs with large potential trade, exchange-rate volatility had a negative and
significant effect on bilateral trade among the countries considered. Dell’Ar-
iccia (1999) examined the effect of exchange-rate volatility on the bilateral
trade of European Union members plus Switzerland over the period 1975–
1994 using several definitions of volatility. In the basic OLS regression,
exchange-rate volatility had a small—but significant—negative impact on
trade; reducing volatility to zero in 1994 would have increased trade by
an amount ranging from 10 to 13 per cent, depending on the measure of
volatility used. Using both fixed and random effects, he found the impact
of volatility was still negative and significant, but smaller in magnitude.
The author found that elimination of exchange rate volatility would have
increased trade by about three and a half per cent in 1994. Rose (2000) also
obtained similar results employing a gravity model. His data set consisted
of 186 countries for the five years 1970, 1975, 1980, 1985, and 1990. In
his benchmark results (without random effects), he found that reducing
volatility by one standard deviation would increase bilateral trade by about
13 per cent. Using random effects also, he found a smaller but still signifi-
cant negative effect; reducing volatility by one-standard deviation would
increase bilateral trade by about four per cent. In general, Rose’s results are
consistent with those of Dell’Ariccia.
Tenreyro (2004), however, cast some doubt on the robustness of Rose’s
results. Using annual data from 1970–1997 on a sample of 104 (developed
and developing) countries, and employing a gravity model that took
endogeneity into account, she found that volatility had an insignificant
effect on trade. Clark et al. (2004) applied the gravity model using a battery
of estimation techniques—including fixed and random effects—to a panel
156 Review of World Economics 2008, Vol. 144 (1)

of 178 International Monetary Fund (IMF) members using every fifth year
from 1975–2000. Using both country- and time-fixed effects, the authors
found a negative and significant impact of exchange-rate volatility on trade;
a one-standard deviation fall in exchange-rate volatility, raised trade by
seven per cent. Allowing for time-varying random effects, however, a nega-
tive relationship was not evident. The authors concluded that, while there is
evidence that increased exchange-rate volatility reduces the volume of trade,
this finding depends on the particular estimation technique employed.

3 Analytical Considerations

The argument that exchange-rate volatility reduces trade typically runs as

follows.6 In a two-country context, consider a firm located in country A
that sells its product in country B (as well as in country A). For simplicity,
suppose that the firm sells in a forward market in each country so that the
firm knows the future price of its product at the time it incurs its costs of
production. However, if there is no futures or forward market for foreign
exchange, the firm has an exchange risk for the future conversion of its
sales revenues from country B into the currency of country A. If the firm is
risk-averse, it is willing to incur some added cost to avoid this risk, so that
the risk, if not hedged, is an implicit cost. In the presence of such a cost, this
reasoning suggests that the firm’s supply price at each quantity of export
sales is higher than in the absence of the risk. For such firms in the aggregate,
the quantity of exports supplied at a given price is smaller with this risk than
without it. The same reasoning applies to firms in country B. If the risk is
present for firms in both countries, the supply curve for exports from each
country to the other is shifted to the left, compared with those that would
exist in the absence of exchange-rate risk. Trade is reduced in a way similar
to that resulting from an increase in transportation costs.
Where there is a forward market for foreign exchange, a discount of
the forward exchange rate in one direction, below the expected future rate,
is a premium in the other direction. Thus, if expectations are similar in
the two countries, such a discount cannot be a deterrent to trade in both
directions. However, the brokerage cost (spread) for forward transactions

6 For recent discussions, see McKenzie (1999) and Clark et al. (2004). The discussion in
the text, draws, in part, and expands on those studies as well as on the studies by Bailey
et al. (1987), Dellas and Zilberfarb (1993), and De Grauwe (2005).
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 157

is generally greater than that for spot transactions in foreign exchange, and
the spread is an increasing function of the variability of the exchange rate.
Hence, the risk can be hedged only at a cost; the existence of forward or
futures markets for foreign exchange does not change the thrust of the above
argument although it reduces its quantitative significance.
The arguments, however, are not all on one side. Consider the following
factors, which suggest that exchange-rate volatility can increase trade:

(i) Exporters may gain knowledge through trade that might help them
anticipate future exchange-rate movements better than can the average
participant in the foreign-exchange market. If so, the profitability of this
knowledge could be used to offset the risk of exchange-rate volatility. If
exporters wish to hedge longer-term investment or other transactions,
rather than use the forward-exchange market, they can borrow and
lend in local currency to offset their other commitments. For example,
a plant in a foreign country can be financed mainly with local capital,
so that investors limit their exchange risk in the basic investment.
(ii) A counterargument of especially great weight is that one must specify
the alternative to exchange-rate volatility. If the volatility is attributable
to fundamental factors’ influencing the exchange rate, intervention
by the authorities to reduce it would be unsustainable and eventu-
ally disruptive. To achieve a reduction of apparent, observed volatility,
authorities would have to intervene with exchange controls or other
restrictions on trade and payments. That intervention could reduce
the volume of trade more than would unrestrained movement of the
exchange rate.
(iii) For countries that hold foreign-currency balances, variability of an
exchange rate does not measure the effect added amounts of that foreign
currency have on the overall riskiness on the firm’s asset portfolio.
The latter risk effect depends on the covariance of an exchange rate
with the prices of the firm’s other assets as well as the own variance
of the exchange rate. In particular, the firm may hold a portfolio of
several foreign currencies, thereby diversifying the risk. If variations in
one currency’s exchange rate against the home currency are negatively
correlated with the variations in others, its variability reduces portfolio
risk, rather than increases it when that currency is added to the portfolio.
In general, variance by itself does not measure the exchange risk.
(iv) If firms can adjust factor inputs in response to movements in the
exchange rate, increased variability may create opportunities to raise
158 Review of World Economics 2008, Vol. 144 (1)

profits. That is, movements in exchange rates represent not only risk,
but also potential reward. If a firm adjusts inputs to both high and low
prices of exports in order to take advantage of profit opportunities when
prices are relatively high, its expected (or average) profits will be higher
the higher is exchange-rate volatility because the firm can sell more
when the price is high and less when the price is low. If risk aversion is
relatively low, the positive effect of greater price volatility on expected
profits may outweigh the negative impact of higher profits stemming
from the uncertainty associated with exchange-rate volatility, and the
firm will produce and export more (Clark et al. 2004: 4; De Grauwe
2005: 69–75). As pointed out by De Grauwe (2005: 73), exporting goods
can be viewed as an option, the value of which, rises when the volatility
of the underlying asset increases; when the exchange rate becomes more
favorable, the firm exercises its option to export.
In the light of the foregoing arguments, the issue of the relationship
between exchange-rate volatility and trade appears to be an empirical ques-
tion. In what follows, we describe the approach taken in this paper.

4 The Model and the Data

Following BTU (1986), the model estimated (with one exception) is of the
following form:

log Xi,t = log a1 + a2 log Yi,t + a3 log RPi,t + a4 log OPt

(1)
+ a5 Vi,t + ei,t ,

where Xi,t is the volume of exports of industrial country i, Yi,t is real GDP
of industrial-country trading-partner nations, RPi,t is a measure of relative
prices of exports of country i to those of its trading partners, OPt represents
real export earnings of oil-exporters, Vi,t is real exchange-rate variability,
and ei,t is a random-error term, and t indicates time. The particular countries
considered in constructing measures of X, Y, RP, and OP, are listed below.
In equation (1), the coefficients are assumed to be constants. This strong
assumption is relaxed in RC estimation. Furthermore, the assumptions
about the relationship between ei,t and the explanatory variables in (1)
should be based on the correct interpretation of ei,t . Such assumptions are
used in RC estimation (Swamy and Tavlas 2001). As discussed in the next
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 159

section, RC estimation does not add an arbitrary error term to a mathemat-

ical equation to obtain an econometric model as is done in (1).
Using quarterly time-series data over the interval 1973:1–1984:3, BTU
(1986) estimated time-series regressions based on the above model (1)
for each of the G-7 economies—Canada, France, Germany, Italy, Japan,
the United Kingdom, and the United States. Using OLS, correcting for
autocorrelation where necessary, and employing two measures of nomi-
nal exchange-rate volatility, they estimated a total of 22 regressions.7 The
authors did not find a single instance of a negative and significant coefficient
on the exchange-rate volatility term in any of their regressions.
This study applies the above model to 12 industrial economies—the
G-7 economies plus those of Ireland, the Netherlands, Norway, Spain,
and Switzerland. The data frequency is quarterly and the sample period
is 1977:1–2003:4. All data come from the International Financial Statistics
(IFS).8 In what follows, we describe these data.
The dependent variables in the estimated equations are the real exports
of the countries considered. There are problems involved in devising proxies
for the independent variables. Theory tells us that income in trading-partner
nations should affect a country’s exports. To construct an income variable
for trading partners, we proceeded as follows. Real GDP data are available
in the IFS for the period covered by this study for the following 11 countries:
Australia, Canada, France, Germany, Italy, Japan, the Netherlands, Spain,
Switzerland, the United Kingdom, and the United States.9 The exporting-
country trading-partner income variable for Ireland and Norway (for which
quarterly GDP data were not available over the entire sample period) was
constructed by converting the real GDP data for the 11 countries listed above
to U.S. dollar terms using year-average 1998 exchange rates and summing the
data for all 11 countries. For each of the ten other countries (i.e., other than
Ireland and Norway) that is the subject of this investigation, the industrial-
country trading-partner income variable is constructed by subtracting the
dollar-denominated real GDP data for the country in question from the
7 Specifically, the authors used the absolute value of the quarter-to-quarter percentage
change in the nominal effective exchange rate as a measure of volatility. The variable was
used in both its current period form and as an eight-period, second-degree polynomial
lag.
8 Data series that were not seasonally adjusted in the IFS were tested for seasonal adjust-
ment using the Census X11 program (multiplicative option); those that displayed season-
ality were seasonally adjusted for use here.
9 The choice of both the particular countries and the sample period was constrained by
data availability.
160 Review of World Economics 2008, Vol. 144 (1)

11-country sum, e.g., the industrial-country trading-partner income vari-

able for Switzerland is the dollar-denominated 11-country aggregate GDP
minus the dollar-denominated real GDP for Switzerland. These series were
employed as our industrial-country trading-partner income variable, Yi,t .
In order to aggregate national GDP series, it was necessary to convert them
to a common currency; we chose the U.S. dollar. However, we wanted to
ensure that our income variables were affected by only changes in real in-
comes in partner nations; we did not want the variables to be affected by the
changing foreign-exchange value of the dollar. Thus, we converted all GDP
data to dollars at a set of fixed exchange rates. We valued trading-partner
income in U.S. dollars at average 1998 exchange rates.
While industrial-country trading partners purchase the bulk of the
exports of the countries under study here, since 1973, the oil-exporting
countries have been major purchasers of the exports of these industrial
countries. Since the oil-exporters tend to be at a different stage of devel-
opment from that of the industrial-country trading partners, however, it
is possible that oil-exporters’ income elasticities of demand (with income
proxied by their export earnings) for imports from industrial countries
differ from the income elasticities of industrial countries. Accordingly, the
oil-exporter “income” (i.e., export earnings) variable enters our export
equations separately from the GDPs of industrial-country trading partners.
The ability of oil exporters to purchase foreign goods varies with the real
purchasing power of their exports rather than the countries’ real outputs,
and export earnings of these countries can vary with the price of oil even
as their real GDPs move in the opposite direction. Thus, the oil-exporter-
income variable is the sum of the dollar values of the oil-exporters’ export
earnings deflated by the dollar-denominated export unit value index of all
industrial countries taken as a whole.10
Theoretically, the relative-price variables in the export equations should
be the ratio of export prices in country i to the domestic prices of similar
goods produced by its trading partners. Since that measure is not available,
the relative-price variable in our export regressions is a real-exchange-rate
index for each country. This variable is based on unit labor costs in manu-
facturing and represents the product of the index of the ratio of the relevant

10 The following countries comprise the IMF’s oil-exporters composite and are used in
constructing the oil-exporter-income variable used in this study: Algeria, Indonesia, Iran,
Iraq, Kuwait, Libya, Nigeria, Oman, Qatar, Saudi Arabia, the United Arab Emirates, and
Venezuela.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 161

indicator of the country considered to a weighted geometric average of the

corresponding indicators for twenty other industrial counties.
Exchange-rate volatility has been measured in the literature using either
nominal or real (effective) exchange rates. As nominal and real exchange
rates tend to move closely together, given the stickiness of domestic prices
(especially in the short run), the choice of measure is not likely to affect the
econometric results. The decision to engage in international transactions,
however, stretches over a relatively long period of time, during which pro-
duction cost and import prices in foreign-currency terms are likely to vary.
This latter consideration suggests that exchange rates measured in real terms
are appropriate, and we have, therefore, used real rates.11 For each country,
three volatility measures are tested. Our first two measures, described below,
were used by BTU (1986) in terms of nominal values.
(1) One measure is the absolute values of the quarterly percentage change
in the exporting nation’s effective exchange rate
Ai,t = |(Ei,t − Ei,t−1 )/Ei,t−1 | , (2)
where Ei,t is the real effective exchange rate of the currency of export-
ing nation i. This measure of volatility is used to test for a stable and
significant response of exports to a one-percentage-point change in the
exchange rate.
(2) A second measure is the following: the log of the eight-quarter moving
standard deviation of the real effective exchange rate

8 1/2
1
Si,t = (Ei,t+k−1 − Ei,t+k−2 )2
. (3)
8 k=1
Both this measure and the previous measure capture delayed responses
of exports to exchange-rate volatility. This second measure is used to
test for a stable and significant response of exports to a one–per cent
change in the standard deviation.
(3) In recent years, some authors have attempted to capture exchange-rate
volatility by using the conditional second moment as a proxy (e.g., Chou
2000; Clark et al. 2004; Siregar and Rajan 2004). The underlying idea is
that part of the volatility can be predicted based on past values of the
exchange rate. Therefore, firms engaged in trade would likely make an

11 In his literature survey, McKenzie (1999: 85) concluded that the distinction between
real and nominal rates has not significantly affected the results derived.
162 Review of World Economics 2008, Vol. 144 (1)

effort to develop such a forecast. We constructed a GARCH measure of

volatility as follows:
Ei,t = α0 + α1 Ei,t−1 + ui,t , (4a)

hi,t = α + βu2i,t−1 + γ hi,t−1 , (4b)

where the exchange rates are expressed in logs and ui,t is a random
error. The conditional variance equation in (4b) is a function of three
terms: (i) the mean, α; (ii) news about volatility from the previous
period, measured as the lag of the squared residual from the mean
equation, u2i,t−1 (the ARCH term); and (iii) the last period’s forecast
error variance, hi,t−1 (the GARCH term).12 We estimated a number
of versions of GARCH models. For equation (4a), lags of up to three
periods were used, depending upon whether the lags were significant.
A GARCH (1,1) specification generated superior results.13

5 Estimation Methods

This section briefly describes the six estimation techniques used.14 We as-
sume, realistically we believe, that RC estimation is likely to be less familiar
to readers than the other approaches used. Therefore, we devote somewhat
more space to describing the RC procedure.
(i) Common-fixed coefficients. This approach applies OLS to the panel
data, allowing the intercept and slopes of (1) to be the same for all
the countries and time periods we considered. Under this assumption,
(1) may not provide an adequate approximation to the “true” model
(Swamy and Tavlas 2001).

12 It follows from the specification of a time-varying coefficient (TVC) model below that
the conditions that the coefficients of (4a) are constant and E[ui,t |Ei,t−s−1 ] = 0 for all
s ≥ 0, may imply misspecifications in (4a) (Swamy and Tavlas 1994; Christou et al. 1998).
13 Siregar and Rajan (2004) obtained similar results in their study of Southeastern
Asian economies. These authors, as well as McKenzie (1999), mention potential prob-
lems involved in ARCH-based measures of exchange-rate volatility; since the exchange-rate
volatility generated prior to the end of the sample period incorporates knowledge about
the future, ARCH models are estimated over the entire sample period. To overcome this
problem one would need to re-estimate the ARCH model at the beginning of each quar-
ter using information that is known to the trader at that time.
14 Baltagi (2001) provides a comprehensive discussion of one-way and two-way fixed and
random effects models and their use in panel-data analyses.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 163

(ii) Fixed effects. Suppose that certain unobserved country-specific variables

that are constant over time t influence the dependent variable of (1) and
are correlated with the explanatory variables in the equation. Under this
assumption, a country-specific constant term is added to the right-hand
side of (1) to allow the equation contain the country-specific variables.15
In this connection, some authors have claimed (e.g., Clark et al. 2004)
that country-specific constant terms help control for remoteness or
multilateral resistance effects. The concept of multilateral resistance
was proposed by Anderson and Van Wincoop (2003), who defined it
as a function of unobservable equilibrium price indices that depend on
bilateral trade barriers and income shares of the trading partners.
(iii) Random effects. If the unobserved country-specific variables repre-
sented by a country-specific constant term are uncorrelated with the
explanatory variables of (1), then the random-effects approach specifies
that the country-specific term is a country-specific random element,
similar to ei,t , except that, for each country, there is but a single draw
that enters (1) identically in each period. With random effects, the error
term has two components: the traditional error unique to each observa-
tion and an error term representing the extent to which the intercept of
the ith country differs from the overall intercept. The composite error
term is nonspherical, so that generalized least squares (GLS) estimation
is used.16
(iv) Generalized method of moments (GMM). Equation (1) is extended to
include log Xi,t−1 , log Xi,t−2 , and logXi,t−3 as additional explanatory
variables. A GMM is used to estimate this extended equation with lagged
independent variables acting as instruments; in the seven-country
panel, five lags of each of the independent variables were used while
in the 12-country panel three lags of each of the independent vari-
ables were used. Since there are more instruments than right-hand side
variables, the estimated regression equations are overidentified. To as-
sess the validity of the different specifications we compute the Sargan
(1964) test for overidentifying restrictions, which amounts to a test of
the exogeneity of the explanatory variables, and AR1 and AR2 tests for
autocorrelation which are listed as m1 and m2 in Arellano and Bond
(1991).
15 Additionally, a time-specific constant term could be introduced.
16 Dell’Ariccia (1999) claims that use of both fixed and random effects can help deal with
simultaneity problems. The Swamy and Arora (1972) estimators of the component vari-
ances are employed in the estimation of the random-effects equations.
164 Review of World Economics 2008, Vol. 144 (1)

Each of the four above estimation techniques imposes assumptions that

can be difficult to fulfill. Common-fixed-coefficients estimation assumes
that the intercept and slopes are the same for all countries and time periods.
We have already noted the serious implications stemming from the assump-
tions made by common-fixed-coefficients estimation. With regard to fixed-
effects estimation, for consistent estimation of (1) using the OLS method,
a necessary condition is that the country-specific variables and the included
explanatory variables in (1) are independent of ei,t . Under random effects,
a necessary condition for the GLS estimator of the coefficients of (1) to be
consistent is that the error components are independent of the explanatory
variables in (1). GMM estimators can be inconsistent because there are
obstacles in obtaining the instrumental variables needed to apply GMM,
as shown by Swamy and Tavlas (2007) in their derivation of an adequate
approximation to the “true” model. In the light of these factors, we turn to
a procedure that can produce consistent estimators of the coefficients.

(v) First-generation random-coefficient (RC) estimation. Effectively, first-

generation RCs capture the effects of nonlinearity through the time
variation of the coefficients.
(vi) Second-generation random-coefficient (RC) estimation. In this estima-
tion, not only the intercept but also the slopes of (1) are allowed to differ
among the countries both at every point in time and through time. In
this form, (1) is referred to as “the time-varying coefficient (TVC) mod-
el”. Let a∗1i,t = log a1i,t + ei,t , a∗2i,t , a∗3i,t , a∗4i,t , and a∗5i,t be the intercept
and the coefficients on log Yi,t , log RPi,t , log OPt , Vi,t , respectively,
in the TVC model. Then it follows from Swamy and Tavlas (2001)
that when the “true” model exists, the TVC model is its exact represen-
tation with unique coefficients if (i) the intercept, a∗1i,t , is interpreted
as the sum of (a) the intercept of the “true” model, (b) the joint effect
on log Xi,t of the portions of excluded variables (i.e., the determinants
of log Xi,t excluded from (1)) remaining after the effects of the “true”
values of the explanatory variables in (1) have been removed, and
(c) the measurement error in log Xi,t , and (ii) for j = 2, 3, 4, 5, a∗ji,t is
interpreted as the sum of (a) the jth coefficient of the “true” model
which is also called “the bias-free component of the jth coefficient
of the TVC model”, (b) a term capturing omitted-variables bias due
to excluded variables, and (c) a measurement-error bias due to mis-
measuring the jth explanatory variable in (1). These are the correct
interpretations of the coefficients of the TVC model.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 165

We estimate the TVC model under the following RC assumption: For

all i = 1, 2, ..., n(= 7 or 12) and t = 1, 2, ..., T(= 108):
a∗i,t = ā∗i + Zi,t π ∗ + ε∗i,t , (5a)

ε∗i,t = Φii ε∗i,t−1 + vi,t

∗
, (5b)
where a∗i,t = (a∗1i,t , a∗2i,t , ..., a∗5i,t ) , ā∗i = (ā∗1i , ā∗2i , ..., ā∗5i ) , Zi,t is a 5 × (p − 1)
matrix of observations on (p − 1) non-constant coefficient drivers, π ∗
is a (p − 1)-vector of fixed coefficients, ε∗i,t = (ε∗1i,t , ε∗2i,t , ..., ε∗5i,t ) , Φi,i is
∗ ∗ ∗ ∗
a 5 × 5 matrix, vi,t = (v1i,t , v2i,t , ..., v5i,t ) , the ā∗i are independently dis-
tributed with mean vector, Eā∗i = ā∗ = (ā∗1 , ā∗2 , ..., ā∗5 ) , and covariance
∗
matrix, ∆, the vi,t are independently distributed with mean zero and co-
variance matrix, ∆i,i , ε∗i,t is independent of ā∗i , Zi,t is independent of ε∗i,t and
ā∗i , and a∗i,t is conditionally independent of the explanatory variables of (1),
given Zi,t . Intuitively, the coefficient drivers can be thought of as variables,
that serve two purposes. First, they deal with the correlation between the
included explanatory variables (log Yi,t , log RPi,t , log OPt , Vi,t ) and their
coefficients in the TVC model. In other words, even though the included
explanatory variables are not unconditionally independent of their coeffi-
cients, they can be conditionally independent of their coefficients given the
coefficient drivers. Second, the coefficient drivers allow us to decompose
the coefficients of the TVC model into their respective components. We call
the estimation of the TVC model under assumptions (5a) and (5b) “the
RC estimation” because under these assumptions, the TVCs are random
variables. This RC estimation is called “first generation” if π ∗ = 0 and is
called “second generation” otherwise.
We may have to include appropriate coefficient drivers with nonzero
coefficients on the right-hand side of (5a) to make the assumption of condi-
tional independence between the included explanatory variables and their
coefficients in the TVC model hold. We decompose the right-hand side
of (5a) into two parts so that for j = 2, 3, 4, 5, one part measures the jth coef-
ficient of the “true” model and the other part measures the sums of omitted-
variable and measurement-error biases contained in a∗ji,t . The measure of
the jth coefficient of the “true” model is an estimate of the jth coefficient
of the TVC model corrected for omitted-variable and measurement-error
biases. When the jth coefficient of the “true” model with j = 2, 3, 4, or 5 is
zero, the correlation between log Xi,t and the jth explanatory variable with
j = 2, 3, 4, or 5 in (1) is spurious (Swamy and Tavlas 2007). We consider
both zero and nonzero values of Φi,i .
166 Review of World Economics 2008, Vol. 144 (1)

6 Empirical Results

The following sets of regressions were estimated. First, the four constant-
coefficient panel-estimation methods were applied to the BTU model—i.e.,
the specification including real export earnings of oil exporters—using data
both for the G-7 countries—Canada, France, Germany, Italy, Japan, the
United Kingdom, and the United States—and for the expanded set of 12
countries. Second, the same four constant-coefficient estimation methods
were applied to the same groups of countries using a specification that
excludes the export earnings of oil exporters.
Third, first-generation RC estimation—i.e, the RC method employing
only the time variation of the coefficients—was applied to the two sets of
countries using both the BTU specification and the specification excluding
export earnings of oil exporters. That is, we did not use coefficient drivers to
decompose the coefficients into their respective components, a procedure
that provides implied estimates of the bias-free components (i.e., the com-
ponents free of omitted-variable and measurement-error biases) of RCs.17
These first-generation RC regressions, are, however, not free of misspeci-
fications because they do not take account of the correlations between the
included explanatory variables and their coefficients. As well, in the RC envi-
ronment the distributional assumptions made about the coefficients in this
first-generation RC procedure can be inconsistent with the “correct” inter-
pretation of the coefficients, whereby, under the correct interpretation, each
of the slope coefficients is the sum of three terms: (1) a bias-free compon-
ent, (2) an omitted-variables bias component, and (3) a measurement-error
bias component. Because each slope coefficient is, in fact, the sum of these
three components, the pattern of variation in each of these components
may be inconsistent with the assumed pattern of variation of the sum. It is,
therefore, important to isolate the bias-free component.
Fourth, to obtain bias-free components, we then estimated regressions
using “coefficient drivers”, which provide second-generation RC results.
As in the case with the first-generation technology, the second-generation
approach was applied to both model specifications (i.e., with and without
export earnings of oil exporters) and to both groups of countries. Two
coefficient drivers were used: z1t = the change in the real exchange rate in
period t − 1 for each country considered, that is, the change (in period t − 1)

17 Nevertheless, first-generation RC sheds light on specification errors due to incorrect

functional forms in fixed-coefficient methods.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 167

of each country’s relative-price variable, denoted as RP in equation (1),

and z2t = the change in exchange-rate volatility in period t − 1, whereby
the exchange-rate volatility measure is defined according to the measure
employed in a particular equation.18
In what follows, specifications with the absolute-per-cent-change mea-
sure of volatility are identified with the subscript “a”, those with the moving-
standard-deviation measure with the subscript “b”, and those using the
GARCH measure with the subscript “c”. For the RC regressions, the average
of the coefficients over the entire time period and all cross sectional units
are reported.
All the reported results for GMM estimation are estimates of the coeffi-
cients of model (1) with large-sample t-ratios. Two types of diagnostic tests
were performed for all the estimated equations. First, the m1 and m2 tests for
autocorrelation were performed. For the BTU model, the following results
were obtained: the values of m1 for the seven-country panel were 4a: 0.15,
4b: 0.04, 4c: 0.09 and for the 12-country panel 4a: 0.55, 4b: 0.60, 4c: 0.67.
The values of m2 for the seven-country panel were 4a: 1.66, 4b: 1.49, 4c: 1.64,
and for the 12-country panel 4a: −1.13, 4b: −0.49, 4c: −0.83. For the model
that omits the variable representing real export earnings of oil exporters,
the following results were obtained: the values of m1 for the seven-country
panel were 4a: −0.06, 4b: −0.10, 4c: 0.07 and for the 12-country panel
4a: −0.70, 4b: −0.34, 4c: −0.02. The values of m2 for the seven-country
panel were 4a: −0.69, 4b: −0.68, 4c: −0.51, and for the 12-country panel
4a: 0.10, 4b: 0.10, 4c: 0.19. None of these values rejects the hypotheses that
there is no serial correlation in the regression disturbances. Next, the Sargan
test was performed. For the BTU model, the following results were obtained:
the values for the seven-country panel were 4a: 27.93, 4b: 22.61, 4c: 24.15
and for the 12-country panel 4a: 10.90, 4b: 8.28, 4c: 10.79. For the model that
omits the variable representing real export earnings of oil exporters, the fol-
lowing values were obtained for the Sargan test for the seven-country panel
4a: 663.10, 4b: 662.50, 4c: 665.20 and for the 12-country panel 4a: 1081.00,
4b: 1165.00, 4c: 1172.00. None of these values rejects the overidentifying
restrictions.
Table 1 presents regression results for the seven-country panel using the
four constant-coefficient panel-data methods. For the sake of brevity, in

18 The bias-free components of the coefficients of second-generation RC models are esti-

∗ ∗
mated using the estimator, āˆ j + Z1i,t π̂j∗ , where āˆ j and π̂j∗ are the iteratively rescaled GLS
∗ ∗
estimators of āj and πj , respectively, for j = 2, 3, 4, 5.
168 Review of World Economics 2008, Vol. 144 (1)

Table 1: Panel Data Estimation of Export Equations: Seven Countries, BTU Model

Estimation Constant Industrial- Oil-exporter Real Exchange- Standard

method country export exchange rate error of
income earnings rate volatility regression
measure

Common fixed coefficients

1a −2.977 0.758 0.218 −0.137 −0.002 0.287575
(−1.64) (3.05) (3.26) (−0.77) (−0.34)
1b −2.891 0.75 0.217 −0.124 −0.022 0.287317
(−1.61) (2.95) (3.36) (−0.63) (−0.64)
1c −2.953 0.757 0.219 −0.141 −1.776 0.287495
(−1.60) (3.06) (3.18) (−0.79) (−0.52)
Fixed effects
2a −10.172 1.662 0.085 −0.275 0.001 0.097147
(−12.79) (18.47) (2.05) (−4.63) (0.68)
2b −10.099 1.652 0.084 −0.267 −0.010 0.097040
(−11.94) (16.61) (2.05) (−4.13) (−0.52)
2c −10.143 1.659 0.085 −0.275 −0.154 0.097175
(−12.55) (18.26) (2.06) (−4.65) (−0.14)
Random effects
3a −9.934 1.625 0.090 −0.257 0.001 0.110488
(−11.42) (18.06) (2.08) (−4.40) (0.50)
3b −9.859 1.615 0.089 −0.249 −0.010 0.110368
(−10.75) (16.29) (2.10) (−3.79) (−0.58)
3c −9.902 1.622 0.090 −0.257 −0.260 0.110697
(−11.24) (17.92) (2.08) (−4.40) (−0.25)
GMM estimation
4a −9.565 1.690 0.074 −0.502 0.001 0.033056
(−6.95) (14.23) (3.22) (−3.80) (0.25)
4b −9.556 1.694 0.073 −0.509 −0.005 0.032864
(−6.97) (13.02) (2.94) (−4.11) (−0.16)
4c −9.512 1.689 0.073 −0.510 −0.373 0.033014
(−6.36) (13.13) (3.17) (−3.72) (−0.05)

Note: The estimation period for all the models is 1977:1–2003:4. The figures in parentheses
are the t-ratios.

Table 1 and in subsequent tables, common-fixed-coefficients estimation is

identified as method 1, fixed effects as method 2, random effects as method 3,
and GMM as method 4. The following results merit comment. First, the
coefficients on the volatility variables are insignificant in each of the 12
regressions reported. Second, the coefficient on the oil-exporter “income”
variable is significant and positive in each equation, and markedly lower
than the coefficient on the industrial-country income variable, indicating
that oil exporters and industrial countries have different income elasticities,
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 169

as hypothesized. Third, comparing the coefficients of the three explana-

tory variables other than volatility (i.e., industrial-country trading-partner
income, oil-exporter income, the real exchange rate) among the four sets
of regressions, there is a clear delineation between the results based on
common-fixed coefficients (method 1) and the results of the other three
methods. Each of the other three methods yields industrial-country income
elasticities in the range of 1.6 to 1.7, more than twice those provided by
method 1. Methods 2 through 4 give oil-exporter-income elasticities in
the range of 0.07 to 0.09, about one-third of the elasticities obtained by
method 1. Also, method 1 yields low relative-price elasticities compared
with the other three methods. Finally, unlike the other methods, method 1
produces coefficients on real exchange rate variables that are insignificant.
Table 2 reports the results of the panel based on 12 countries. Again, there
is no evidence that exchange-rate volatility reduces trade; the coefficient of
the volatility term is insignificant in each of the 12 equations reported.
As was the case with the results based on the seven-country panel, the
common-fixed-coefficients method gives lower elasticities for industrial-
country income and higher elasticities for oil-exporter income than does
each of the other three methods. Compared with the results reported in
Table 1, the results of methods 2 through 4 (in Table 2) generally show higher
income (both industrial country and oil exporter) and higher absolute
relative-price elasticities.
To what extent do the results reported in Tables 1 and 2 reflect the use
of the BTU specification, which, unlike most specifications found in the
literature, includes real export earnings of oil exporters? To shed light on
this issue, in the context of the fixed-coefficient estimation methods, we re-
estimated each of the specifications contained in Table 1 (seven countries)
and Table 2 (twelve countries), but with the variable capturing the real
export earnings of oil exporters dropped. In what follows, we refer to this
specification as the “standard” specification, since it corresponds to the basic
model typically used in the literature.19 The results are reported in Tables 3
(seven countries) and 4 (twelve countries). Among the 24 specifications
contained in Tables 3 and 4, there is evidence of a negative and significant
impact of volatility under two specifications—those denoted as 4c in each
table—that is, the specification estimated with GMM and using the GARCH
measure of volatility. This result is consistent with that found in some of
the recent literature, which shows some tendency to report a significant and

19 Of course, the basic specification is often augmented with additional variables.

170 Review of World Economics 2008, Vol. 144 (1)

Table 2: Panel Data Estimation of Export Equations: Twelve Countries, BTU Model

Estimation Constant Industrial- Oil-exporter Real Exchange- Standard

method country export exchange rate error of
income earnings rate volatility regression
measure

Common fixed coefficients

1a −5.564 1.238 0.195 −0.480 −0.011 0.914058
(−1.36) (2.87) (2.84) (−1.18) (−1.18)
1b −5.212 1.174 0.189 −0.397 −0.110 0.912097
(−1.32) (2.91) (3.03) (−1.00) (−1.13)
1c −5.515 1.236 0.200 −0.493 −9.040 0.913945
(−1.31) (2.84) (2.89) (−1.19) (−0.84)
Fixed effects
2a −10.779 1.869 0.106 −0.553 0.001 0.139874
(−8.21) (15.08) (3.55) (−2.76) (0.20)
2b −10.649 1.851 0.105 −0.537 −0.019 0.139484
(−8.57) (15.71) (3.61) (−2.52) (−0.67)
2c −10.764 1.868 0.106 −0.553 −0.115 0.139878
(−8.05) (14.96) (3.54) (−2.76) (−0.12)
Random effects
3a −10.778 1.869 0.107 −0.552 0.001 0.139760
(−9.57) (15.14) (3.56) (−2.77) (0.20)
3b −10.647 1.851 0.105 −0.537 −0.019 0.139369
(−9.89) (15.77) (3.63) (−2.53) (−0.68)
3c −10.763 1.868 0.106 −0.553 −0.116 0.139762
(−9.37) (15.02) (3.55) (−2.77) (−0.12)
GMM estimation
4a −10.623 1.858 0.089 −0.629 0.001 0.051349
(−7.54) (14.96) (2.62) (−3.47) (0.50)
4b −10.586 1.856 0.091 −0.632 0.002 0.054039
(−7.74) (15.30) (2.59) (−3.27) (0.08)
4c −10.549 1.850 0.091 −0.630 −2.261 0.051411
(−7.25) (14.50) (2.71) (−3.45) (−1.02)

Note: The estimation period for all the models is 1977:1–2003:4. The figures in parentheses
are the t-ratios.

negative impact of volatility when using a specification that: (1) includes

a GARCH measure of volatility, and (2) is estimated using GMM (McKenzie
1999; Clark et al. 2004).
Comparing the results in Tables 1 and 2 with those contained in Table 3
and 4, respectively, the following findings are worth singling-out. First, the
coefficients on the variables representing industrial-country income and the
real exchange rate are little changed when dropping the variable representing
real export earning of oil exporters. Second, the standard errors of the
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 171
Table 3: Panel Data Estimation of Export Equations: Seven Countries,
Standard Model
Estimation Constant Industrial- Oil-exporter Real Exchange- Standard
method country export exchange rate error of
income earnings rate volatility regression
measure

Common fixed coefficients

1a −2.310 0.786 — −0.136 −0.003 0.295372
(−1.14) (3.13) (−0.72) (−0.45)
1b −2.224 0.772 — −0.120 −0.026 0.295050
(−1.10) (2.98) (−0.57) (−0.66)
1c −2.319 0.787 — −0.138 −1.531 0.295388
(−1.13) (3.15) (−0.72) (−0.45)
Fixed effects
2a −10.015 1.686 — −0.280 0.001 0.100578
(−12.66) (18.21) (−4.32) (0.41)
2b −9.942 1.676 — −0.271 −0.011 0.100408
(−11.97) (16.65) (−3.96) (−0.58)
2c −9.999 1.685 — −0.280 −0.018 0.100590
(−12.42) (18.00) (−4.32) (−0.02)
Random effects
3a −9.746 1.648 — −0.261 0.001 0.114776
(−11.03) (18.13) (−4.02) (0.23)
3b −9.672 1.637 — −0.251 −0.012 0.114580
(−10.50) (16.50) (−3.49) (−0.64)
3c −9.727 1.646 — −0.261 −0.124 0.114986
(−10.87) (18.01) (−4.01) (−0.11)
GMM estimation
4a −9.031 1.682 — −0.511 −0.001 0.035255
(−7.53) (14.57) (−3.43) (−0.33)
4b −9.066 1.687 — −0.516 0.003 0.035261
(−6.82) (12.40) (−3.21) (0.07)
4c −8.684 1.650 — −0.519 −9.386 0.035102
(−6.11) (13.45) (−3.28) (−2.14)

Note: The estimation period for all the models is 1977:1–2003:4. The figures in parentheses
are the t-ratios.

regressions reported in Tables 1 and 2 are slightly, but uniformly, lower than
those reported in Tables 3 and 4. Thus, while the variable representing real
export earnings of oil exporters does not have much of an impact on the
coefficients of the other included variables, there is some evidence (in terms
of the differences in the values of the coefficients on the industrial-country-
income and the oil-exporter-export-earnings variables) that it should be
included in export equations.
172 Review of World Economics 2008, Vol. 144 (1)

Table 4: Panel Data Estimation of Export Equations: Twelve Countries,

Standard Model
Estimation Constant Industrial- Oil-exporter Real Exchange- Standard
method country export exchange rate error of
income earnings rate volatility regression
measure

Common fixed coefficients

1a −4.652 1.241 — −0.506 −0.010 0.910664
(−1.07) (2.83) (−1.21) (−1.22)
1b −4.684 1.202 — −0.393 −0.114 0.913653
(−1.13) (2.98) (−0.97) (−1.15)
1c −4.991 1.269 — −0.490 −8.702 0.915724
(−1.12) (2.92) (−1.16) (−0.81)
Fixed effects
2a −10.644 1.897 — −0.541 0.001 0.140838
(−8.28) (15.06) (−2.65) (0.05)
2b −10.498 1.880 — −0.533 −0.021 0.143196
(−8.30) (15.58) (−2.41) (−0.74)
2c −10.631 1.900 — −0.549 0.121 0.143660
(−7.78) (14.77) (−2.64) (0.11)
Random effects
3a −10.778 1.867 — −0.541 0.001 0.140720
(−9.57) (15.12) (−2.66) (0.05)
3b −10.496 1.880 — −0.533 −0.021 0.143080
(−9.53) (15.64) (−2.42) (−0.75)
3c −10.630 1.900 — −0.549 0.119 0.143543
(−9.02) (14.82) (−2.65) (0.11)
GMM estimation
4a −8.611 1.839 — −0.920 0.011 0.038999
(−5.42) (13.23) (−6.30) (1.67)
4b −8.510 1.874 — −0.985 0.058 0.038961
(−5.24) (11.15) (−5.94) (1.06)
4c −7.781 1.783 — −0.942 −12.557 0.038916
(−4.56) (11.55) (−6.30) (−2.33)

Note: The estimation period for all the models is 1977:1–2003:4. The figures in parentheses
are the t-ratios.

At this point, a brief summary of the above findings may be helpful

in setting the stage for what follows. Using the BTU specifications, we
found no evidence of an impact of exchange-rate volatility on exports
under any of the four constant-coefficient panel-data methods employed.
When we dropped the variable representing real export earnings of
oil exporters, we found evidence of a significant and negative impact
of volatility in only two specifications, both using GMM estimation and
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 173

a GARCH measure of volatility. How do we discriminate among these

specifications?
To address this issue, we turn to RC estimation, which, as noted, cor-
rects for the specification biases contained in fixed-coefficient procedures.
We apply RC estimation in two steps. First, we use first-generation RC esti-
mation, which, although it does not correct for biases attributable to omitted
variables and measurement errors, does correct for functional-form mis-
specifications found in fixed-coefficient procedures. Second, we then apply
second-generation RC estimation, which corrects for biases attributable to
incorrect functional forms, omitted variables, and measurement errors. The
two RC procedures are applied to both data sets—i.e., to the seven-country
and 12-country panels—and to both models—i.e., the BTU and standard
models.
Tables 5 and 6 report RC results for the seven-country panel and the
12-country panel, respectively. Each table consists of 12 rows. Rows 1
through 3 give results for first-generation RC applied to the BTU model,
while rows 4 through 6 give results for first-generation RC applied to the
standard model. Rows 7 through 9 provide results for second-generation RC
applied to the BTU model, while rows 10 through 12 give second-generation
RC applied to the standard model. In these tables, first-generation specifica-
tions are denoted as RC1 and second-generation specifications are denoted
as RC2. As in the previous tables, subscripts “a”, “b”, and “c” refer to the
volatility measures used.
The following results contained in Table 5 (seven-country panel) are
worth noting. First, in comparing the BTU model with the standard model
using first-generation RC technology (i.e., rows 1–3 and 4–6, respectively),
there is no evidence of a significant volatility effect.20 In fact, the high-
est absolute value of the t-ratios is provided by the GARCH measure
of volatility in the standard model (row 6), but it is positive; at 1.63,
it is not significant at the 10 per cent level. Second, similar conclusions
apply using second-generation RC technology; the highest absolute value of
20 Nevertheless, the RC approach does not drop a variable under the condition that its
coefficient is insignificant. A full treatment of RC estimation assesses whether the inclu-
sion of a volatility measure in (1) could be reducing omitted-variable and measurement-
error biases contained in the coefficients of the equation compared to what they would
have been in the absence of a volatility measure. In other words, the fact that a volatil-
ity measure is insignificant and contributes little to the coefficient of determination does
not in itself provide grounds for excluding the variable. The conditions needed to pursue
this line of research are very difficult to implement in RC estimation using panel data and
we leave this research for a future line of work.
174 Review of World Economics 2008, Vol. 144 (1)

Table 5: Panel Data Estimation of Export Equations Using RC Estimation:

Seven Countries
Estimation Constant Industrial- Oil-exporter Real Exchange- Standard
method country export exchange rate error of
income earnings rate volatility regression
measure

First-generation RC: BTU model

1a (RC1) −9.559 1.627 0.082 −0.336 −0.001 0.000324
(−12.33) (16.48) (1.98) (−3.36) (−0.35)
2b (RC1) −9.211 1.604 0.080 −0.356 −0.014 0.000332
(−10.00) (14.88) (2.11) (−3.19) (−1.05)
3c (RC1) −10.074 1.659 0.084 −0.294 −13.394 0.000322
(−13.01) (14.65) (1.91) (−3.20) (−0.59)
First-generation RC: standard model
4a (RC1) −9.139 1.632 — −0.357 −0.001 0.001074
(−9.28) (14.82) (−2.85) (−0.61)
5b (RC1) −8.863 1.607 — −0.361 −0.026 0.000655
(−7.31) (13.07) (−2.58) (−1.52)
6c (RC1) −9.791 1.704 — −0.370 23.130 0.000554
(−8.41) (12.87) (−3.36) (1.63)
Second-generation RC: BTU model
7a (RC2) −9.435 1.629 0.079 −0.361 −0.001 0.000310
(−10.96) (15.85) (2.01) (−3.35) (−0.08)
8b (RC2) −9.040 1.603 0.077 −0.387 −0.027 0.000210
(−6.98) (12.46) (2.56) (−2.77) (−0.85)
9c (RC2) −9.956 1.651 0.083 −0.301 10.667 0.000043
(−15.10) (16.05) (1.95) (−3.34) (0.94)
Second-generation RC: standard model
10a (RC2) −9.041 1.632 — −0.380 −0.001 0.000760
(−7.96) (14.08) (−2.73) (−0.28)
11b (RC2) −8.716 1.603 — −0.384 −0.027 0.000561
(−5.47) (11.23) (−2.27) (−1.30)
12c (RC2) −9.816 1.717 — −0.380 −21.377 0.000356
(−11.81) (13.02) (−3.12) (−1.43)

Note: The estimation period for all the models is 1977:1–2003:4. In the second-generation
RC model two coefficient drivers are used: change in the real exchange rate in period t − 1
and change in exchange-rate volatility measure in period t − 1. The estimated coefficients of
second-generation RC models are estimated using the change in real exchange rate in the pre-
vious period. The figures in parentheses are the t-ratios.

the t-ratios on volatility is obtained using the standard model and the
GARCH volatility measure; the t-ratio is −1.43, but is not significant.
Third, the coefficients on real income, real export earnings, and the real-
exchange-rate variable are similar to the results reported in the previous
tables.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 175
Table 6: Panel Data Estimation of Export Equations Using RC Estimation:
Twelve Countries
Estimation Constant Industrial- Oil-exporter Real Exchange- Standard
method country export exchange rate error of
income earnings rate volatility regression
measure

First-generation RC: BTU model

1a (RC1) −11.497 1.840 0.101 −0.337 −0.001 0.000359
(−10.13) (13.76) (3.60) (−3.04) (−0.23)
2b (RC1) −11.415 1.826 0.099 −0.321 −0.012 0.000361
(−9.22) (13.34) (3.90) (−2.91) (−0.82)
3c (RC1) −11.802 1.867 0.103 −0.326 −7.019 0.000387
(−11.08) (13.89) (3.38) (−3.38) (−0.32)
First-generation RC: standard model
4a (RC1) −11.334 1.844 — −0.287 −0.001 0.001282
(−8.79) (13.72) (−2.34) (−0.66)
5b (RC1) −11.188 1.820 — −0.265 −0.022 0.000817
(−8.23) (13.55) (−2.02) (−1.83)
6c (RC1) −11.757 1.884 — −0.277 −4.048 0.000650
(−9.12) (13.15) (−2.45) (−0.21)
Second-generation RC: BTU model
7a (RC2) −11.441 1.842 0.100 −0.353 −0.001 0.000306
(−9.44) (13.59) (3.62) (−2.96) (−0.04)
8b (RC2) −11.308 1.824 0.099 −0.341 −0.013 0.000283
(−8.05) (12.98) (4.49) (−2.58) (−1.11)
9c (RC2) −11.702 1.858 0.100 −0.328 −8.651 0.000038
(−9.57) (12.44) (3.53) (−2.62) (−0.46)
Second-generation RC: standard model
10a (RC2) −11.315 1.847 — −0.297 −0.001 0.000930
(−8.07) (13.47) (−2.22) (−0.56)
11b (RC2) −11.145 1.822 — −0.278 −0.023 0.007466
(−7.08) (13.17) (−1.77) (−1.63)
12c (RC2) −11.683 1.882 — −0.288 −5.256 0.000370
(−8.05) (11.60) (−2.35) (−0.26)

Note: The estimation period for all the models is 1977:1–2003:4. In the second-generation
RC model two coefficient drivers are used: change in the real exchange rate in period t − 1
and change in exchange-rate volatility measure in period t − 1. The estimated coefficients of
second-generation RC models are estimated using the change in real exchange rate in the pre-
vious period. The figures in parentheses are the t-ratios.

In Table 6, there is one measure of the coefficient on exchange-rate

volatility that is negative and significant at the ten per cent level (though
not at the five per cent level). The particular measure is the eight-quarter
moving standard deviation in the standard model using first-generation RC
technology, as reported in row 5b of the table. Recall, however, that first-
176 Review of World Economics 2008, Vol. 144 (1)

generation RC estimation does not correct for omitted-variables bias and/or

measurement-errors bias. Consequently, the foregoing result on volatility
could reflect specification biases stemming from omitted variables and/or
measurement errors. To address this issue, consider the second-generation
RC result applied to the same specification—that is, the standard model
using the eight-quarter moving standard deviation measure of volatility. As
reported in row 11b, the coefficient on the measure of volatility remains
negative and the absolute value of the t-ratio falls to 1.63 (compared with
1.83 using first-generation RC technology); the 1.63 t-ratio is not signifi-
cant at the ten per cent level. Thus, the RC results provide some evidence
of specification biases contained in coefficients that are not corrected for
omitted variables and/or measurement errors.
Several other features, pertaining to the standard errors of the regres-
sions (SERs), of Tables 1 though 6 are worth mentioning. (See the final
columns of the tables.) First, there is a marked tendency in each of the
first four tables for the SERs to fall sharply in moving from common-fixed-
coefficients estimation to either fixed effects or random effects. Second,
GMM estimation provides much lower SERs than any of the other constant-
coefficient methods. Third, the RC methods provide much smaller SERs
than GMM.

7 Concluding Remarks

As discussed above, some recent studies, using panel data, found evidence,
but by no means overwhelming, of a significant and negative impact of
exchange-rate volatility on trade. Although it is difficult to draw gener-
alizations from this finding, two factors seem to be of importance. First,
as noted by McKenzie (1999) in his literature survey, the use of a GARCH
specification—typically GARCH (1,1)—of volatility seems to produce
effects of volatility on trade that are more-consistently negative and signifi-
cant than other specifications. Second, studies employing panel data tended
to find negative and significant effects of volatility on trade, regardless of
the measure of volatility employed.
This paper has investigated the following issue: in the light of the
wide diversity of specifications and accompanying results contained in the
literature concerning a relationship between exchange-rate volatility and
trade, how can a satisfactory or adequate specification be determined. Our
approach in dealing with this issue proceeded as follows.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 177

We began by estimating the BTU model, a specification applied in the

literature some twenty years ago to seven large industrial countries. The
distinctive feature of this model is the inclusion of real export earnings of
oil-exporting economies. BTU’s data set was updated and expanded to also
include five additional industrial countries. The BTU model was generalized
to include a GARCH measure of exchange-rate volatility. It was estimated
using four constant-coefficient panel-data estimation methods. We did not
find any evidence of a negative and significant impact of volatility on trade
using panel-data sets consisting of seven and twelve economies.
We then carried out a further investigation whereby we dropped the vari-
able representing real export earnings of oil exporters, providing
a “standard” model of the relationship between exports and volatility. We
found evidence of a negative and significant impact of volatility on trade
when estimating with GMM and using the GARCH measure of volatility.
This result applied for both the seven-country and 12-country panel-data
sets.
To help discriminate between the BTU specification and the standard
specification, we investigated the extent to which the differing results could
be attributable to specification errors due to the estimation procedures used.
To this end, we began by using the two panel-data sets, estimating both the
BTU and standard specifications, applying first-generation RC technology.
By so doing, the results were corrected for possible functional-form mis-
specification, but not for specification biases due to omitted variables or
measurement errors. We found one instance of a negative and significant
(at the ten per cent level) impact of volatility on exports.
We then applied second-generation RC technology. This procedure cor-
rects for specification biases due to incorrect functional forms, omitted
variables and measurement errors. There was no evidence of a negative and
significant impact of volatility on exports.
The following conclusion emerges from our investigation. The finding
of negative and significant effect of volatility on trade appears to arise from
omitted-variable biases and/or measurement-error biases. This conclusion
follows from both (i) a comparison of the BTU specification and the stan-
dard specification, which omits real export earnings of oil exporters, using
fixed-coefficient panel-data estimation methods, and (ii) from the results
of second-generation RC technology, which takes account of specification
biases.
The implications of our results for future research include the following.
First, to investigate the robustness of our results, the application of second-
178 Review of World Economics 2008, Vol. 144 (1)

generation technology to alternate models used in the literature, including

gravity specifications, would appear to be a worthwhile line of inquiry.
Second, our study focused on the larger industrial countries. It might be
interesting to investigate whether our results hold-up to groups of smaller
industrial countries and/or developing countries, particularly as the latter
economies often lack well-developed financial markets that can provide
hedging instruments against exchange-rate risk.

References

Anderson, J. E., and E. van Wincoop (2003). Gravity with Gravitas: A Solution to
the Border Puzzle. American Economic Review 93 (1): 170–192.
Arellano, M., and S. R. Bond (1991). Some Tests of Specification for Panel Data:
Monte Carlo Evidence and an Application to Employment Equations. Review
of Economic Studies 58 (2): 277–297.
Bailey, M. J., G. S. Tavlas, and M. Ulan (1986). Exchange Rate Variability
and Trade Performance: Evidence for the Big Seven Industrial Countries.
Weltwirtschaftliches Archiv/Review of World Economics 122 (3): 466–477.
Bailey, M. J., G. S. Tavlas, and M. Ulan (1987). The Impact of Exchange Rate
Volatility on Export Growth: Some Theoretical Considerations and Empirical
Results. Journal of Policy Modeling 9 (1): 225–243.
Baltagi, B. (2001). Econometric Analysis of Panel Data. Second Edition. Chichester:
Wiley.
Chang, I., P. A. V. B. Swamy, C. Hallahan, and G. S. Tavlas (2000). A Computa-
tional Approach to Finding Causal Economic Laws. Computational Economics
16 (1–2): 105–136.
Chou, W. (2000). Exchange Rate Variability and China’s Exports. Journal of Com-
parative Economics 28: 61–79.
Christou, C., P. A. V. B. Swamy, and G. S. Tavlas (1998). A General Framework for
Predicting Returns from Multiple Currency Investments. Journal of Economic
Dynamics & Control 22 (7): 977–1000.
Clark, P., N. Tamirisa, and S. J. Wei (2004). A New Look at Exchange Rate Volatil-
ity and Trade Flows. Occasional Papers 235. International Monetary Fund,
Washington, D.C.
Coric, B,. and G. Pugh (2006). The Effects of Exchange Rate Variability on In-
ternational Trade: A Meta Regression Analysis. IESR Working Paper 01-2006.
Institute for Environment and Sustainability Research, Staffordshire University,
UK.
De Grauwe, P. (2005). The Economics of Monetary Union. Sixth Revised Edition.
Oxford: Oxford University Press.
 Hondroyiannis/Swamy/Tavlas/Ulan: Exchange-Rate Volatility and Exports 179

Dell’Ariccia, G. (1999). Exchange Rate Fluctuations and Trade Flows: Evidence

from the European Union. IMF Staff Papers 46 (3): 315–334.
Dellas, H., and B. Z. Zilberfarb (1993). Real Exchange Rate Volatility and Inter-
national Trade: A Re-examination of the Theory. Southern Economic Journal 59
(4): 641–647.
Edison, H. J., and M. Melvin (1990). The Determinants and Implications of the
Choice of an Exchange Rate System. In W. S. Haraf and T. D. Willett (eds.),
Monetary Policy for a Volatile Global Economy. Washington, D.C.: AEI Press.
Greene, W. H. (2003). Econometric Analysis. Fifth Edition. Upper Saddle River,
New Jersey: Prentice Hall.
Hansen, L. P. (1982). Large Sample Properties of Generalized Method of Moments
Estimators. Econometrica 50 (4): 1029–1054.
IMF (International Monetary Fund Research Department) (1984). Exchange Rate
Volatility and World Trade. Occasional Papers 28/7/84. International Monetary
Fund, Washington, D.C.
McKenzie, M. D. (1999). The Impact of Exchange Rate Volatility on International
Trade Flows. Journal of Economic Surveys 13 (1): 71–106.
Rose, A. (2000). One Money, One Market: Estimating the Effect of Common Cur-
rencies on Trade. Economic Policy 15 (30): 7–46.
Sargan, J. D. (1964). Wages and Prices in the United Kingdom: A Study in Econo-
metric Methodology. In P. E. Hart, G. Mills, and J. K. Whitaker (eds.), Econo-
metric Analysis for National Economic Planning. London: Batterworths.
Siregar, R., and R. S. Rajan (2004). Impact of Exchange Rate Volatility on Indone-
sia’s Trade Performance in the 1990s. Journal of the Japanese and International
Economies 18 (2): 218–240.
Swamy, P. A. V. B., and S. S. Arora (1972). The Exact Finite Sample Properties of
the Estimators of Coefficients in the Error Components Model. Econometrica
40 (2): 261–275.
Swamy, P. A. V. B., and G. S. Tavlas (1994). Connections between GARCH and
Stochastic Coefficients (SC) Models. Economics Letters 46 (1): 7–10.
Swamy, P. A. V. B., and G. S. Tavlas (2001). Random Coefficient Models. In
B. H. Baltagi (ed.), A Companion to Theoretical Econometrics. Malden, MA:
Blackwell.
Swamy, P. A. V. B., and G. S. Tavlas (2005). Theoretical Conditions under Which
Monetary Policies Are Effective and Practical Obstacles to Their Verification.
Economic Theory 25 (4): 999–1005.
Swamy, P. A. V. B., and G. S. Tavlas (2006). A Note on Muth’s Rational Expec-
tations Hypothesis: A Time-Varying Coefficient Interpretation. Macroeconomic
Dynamics 10 (3): 415–425.
Swamy, P. A. V. B., and G. S. Tavlas (2007). The New Keynesian Phillips Curve
and Inflation Expectations: Re-specification and Interpretation. Economic
Theory 31 (2): 293–306.
180 Review of World Economics 2008, Vol. 144 (1)

Swamy, P. A. V. B., and P. A. Tinsley (1980). Linear Prediction and Estimation

Methods for Regression Models with Stationary Stochastic Coefficients. Journal
of Econometrics 12 (2): 103–142.
Swamy, P. A. V. B., I. Chang, J. S. Mehta, and G. S. Tavlas (2003). Correcting for
Omitted-Variable and Measurement-Error Bias in Autoregressive Model Esti-
mation with Panel Data. Computational Economics 22 (2): 225–253.
Tenreyro, S. (2004). On the Trade Impact of Nominal Exchange Rate Volatility.
Federal Reserve Bank of Boston, unpublished.
Verbeek, M. (2004). A Guide to Modern Econometrics. Second Edition. Chichester:
Wiley.
Wei, S. J. (1999). Currency Hedging and Goods Trade. European Economic Review
43 (7): 1371–1394.
Zellner, A. (1988). Causality and Causal Laws in Economics. Journal of Economet-
rics 39 (1–2): 7–21.