1
The Impact of Exchange Rate on FDI and The Interdependence of FDI Over Time
Joseph D. ALBA*, Peiming WANG and Donghyun PARK, Nanyang Technological University, Singapore
Abstract We examine the impact of exchange rates on FDI inflows into the United States in the context of a model that allows for interdependence of FDI over time. Interdependence is modeled as a two-state Markov process where the two states can be interpreted as either a favorable or an unfavorable environment for FDI in an industry. We use unbalanced industry-level panel data from the US wholesale trade sector and our analysis yields two main results. First, we find evidence that FDI is interdependent over time. Second, under a favorable FDI environment, the exchange rate has a positive and significant effect on the average rate of FDI inflows. Keywords: Exchange rate, FDI, Markov, unbalanced panel JEL Codes: F31, F21, F23 _______________ * Corresponding author: Joseph D. Alba, Room No. S3-B2B-56, Economics Division, School of Humanities and Social Sciences, Nanyang Technological University (NTU), Singapore 639798 [E-mail] [email protected] [Telephone] (65)6790-6234 [Fax] (65)6792-4217 Co-Author: Peiming Wang, Room No. S3-B1A-33, Banking and Finance Division, Nanyang Business School, Nanyang Technological University (NTU), Singapore 639798 Co-Author: Donghyun Park, Room No. S3-B1A-10, Economics Division, School of Humanities and Social Sciences, Nanyang Technological University (NTU), Singapore 639798
2
1 Introduction
Foreign direct investment (FDI) flows into the Unites States have shown substantial
fluctuations in the 1980s and 1990s. A growing theoretical and empirical literature attempts to
explain those fluctuations primarily in terms of the impact of the real exchange rate on FDI.
Contributions to this literature include Froot and Stein (1991), Blonigen (1997), Klein and
Rosengren (1994), Guo and Trivedi (2002) and Kiyota and Urata (2004). Theoretical
considerations based on the relative wealth and relative labor cost effects suggest that a stronger
US dollar may deter FDI into the US.1 At the same time, however, a stronger US dollar may
improve the home-currency revenues and thus profitability of foreign firms entering the US
market. This helps to explain the entry of foreign firms into the US market during the first half of
the 1980s, when the US dollar appreciated sharply.
Interestingly, there was a tendency among foreign firms to remain in the US market when the
US dollar returned to its original level. Such behavior is an example of hysteresis, or an effect
that persists after its underlying cause has been removed. One possible explanation for the failure
of foreign firms to exit the US market in the face of a falling dollar is the presence of sunk costs
that cannot be recovered upon exit.2 The exchange rate would have to fall below the entry-
triggering level in order to trigger exit. Dixit (1989) further develops the concept of hysteresis by
applying the theory of option pricing from financial economics to analyze investment under
uncertainty.3 Dixit shows that greater price volatility leads to a wider range of prices in which
inactive firms do not enter and active firms do not exit. That is, uncertainty expands the gap
between the entry-triggering price and exit-triggering price, thereby deterring both entry and exit.
Campa (1993) develops an empirically testable model of FDI based on Dixit’s model. Campa’s
model describes a risk-neutral foreign firm that has to incur a sunk cost in order to enter the US
market. It has to decide, at each point in time, whether to enter the US market in this period or
wait until next period. The firm produces a good abroad and can sell it in the US market at a
3
constant dollar price. Although the firm faces a certain price in US dollars, its returns in its home
currency fluctuate if the bilateral exchange rate fluctuates. If the exchange rate is defined as units
of foreign currency per US dollar, a higher exchange rate increases the home currency-profits. At
the same time, the more volatile the exchange rate, the more volatile will be the home-currency
returns, and the wider is the range of exchange rates in which neither entry nor exit occurs.
Campa’s model thus clearly predicts a positive effect of exchange rate and a negative effect of
exchange rate volatility on FDI.4
Campa empirically tests his model using data consisting of a panel based on 61 four-digit
Standard Industrial Classification (SIC) industries in the US wholesale trade sector for the period
from 1981 to 1987. The choice of wholesale industries eliminates the complications of
manufacturing industries pertaining to input origin or final output destination.5 The dependent
variable is the number of foreign firms that entered a US industry in a given year while the
independent variables are measures of exchange rate level R, rate of change in the exchange rate
μ, volatility of the exchange rate σ, sunk costs k, and variable costs of production in the United
States relative to foreign countries w.6 Campa uses a Tobit model to estimate the probability that
an FDI entry occurs in the US wholesale trade sector. The model predicts the probability of entry
is positively related to R and μ, and negatively related to σ, k and w. All variables other than μ
have the predicted sign. Most importantly, the exchange rate level R has a significant positive
effect and the standard deviation of the exchange rate σ has a significant negative effect.7
Tomlin (2000) extends Campa’s sample period to 1993 and uses a zero-inflated Poisson (ZIP)
model to analyze FDI in the US wholesale trade industry. While Campa calculates the probability
that an FDI entry occurs, Tomlin estimates the average rate of FDI entries per industry for the
period 1982 to 1993. Tomlin pools industry data for a period of 12 years, so that her model is in
effect a cross-sectional model that does not consider interdependence over time. In contrast to
Campa, Tomlin finds that neither the level nor the standard deviation of the exchange rate has any
4
effect on the rate of FDI. This suggests that while exchange rate variables may affect the
probability of entry, they do not affect the average rate of FDI entries.
All previous studies of FDI have failed to consider the interdependence of FDI over time. This
possibility is articulated by Caves (1971) using the concept of corporate rivalry in FDI. Caves
argues that rival firms in an oligopoly with product differentiation tend to follow each other in
making direct investments in foreign countries.8 For example, a foreign firm may find that among
its potential markets, an industry in the United States may have a favorable investment
environment. The foreign firm may then decide to enter the US industry. As the first foreign firm
enters the US market, rival firms may also find the investment environment of the US industry
favorable and follow suit. The opposite may happen if a foreign firm finds a better investment
environment in markets outside the United States. A foreign firm may then find the US industry
to be unfavorable to FDI and instead consider other markets. Rival firms may also find the
investment environment in the United States to be unfavorable. Hence, rival firms may view an
industry as favorable or unfavorable to FDI depending on whether their competitors viewed an
industry as favorable or unfavorable to FDI in the previous period.
In the context of corporate rivalry in FDI, whether a foreign firm finds the investment
environment of a US industry favorable or unfavorable may depend not only on the investment
environment in the United States but also on other factors such as its home investment
environment, its interactions with its rivals in markets outside the United States and political
actions of governments affecting it but not its rivals. Since these factors include interactions
among foreign firms and governments as well as changing conditions in various markets, they
may be difficult to measure and subject to uncertainty. Hence, it is impractical to include these
factors as regressors in a model that explains FDI.9
The central focus of our paper is to re-examine the relationship between the exchange rate and
FDI taking into account the possible interdependence of FDI over time, which is described by the
5
Markov zero-inflated Poisson (MZIP) model developed by Wang (2001). More specifically, we
model the interdependence of FDI over time as a two-state Markov process in which the two
states can be interpreted as either a favorable or an unfavorable environment for FDI in an
industry in the United States. The Markov process incorporates the factors affecting the two states
which are difficult to measure and subject to uncertainty. Furthermore, since the MZIP model is
first proposed for a time-series specification but we use industry-level panel data for our
empirical analysis, we formally re-define the MZIP model for panel data and conduct a
simulation study on the applicability of the MZIP model for the analysis of panel data.
Significantly, we address the reclassification of four-digit SIC industry codes after 1987 by
constructing an unbalanced panel data set - i.e. the number of industries in our sample is greater
during 1988-1994 than 1982-1987. We use as our basic empirical framework Campa’s (1993)
empirical application of the theoretical model developed by Dixit (1989). Our results clearly
show evidence of the interdependence of FDI over time as described by the two-state Markov
process and, most critically, our findings empirically re-confirm a significant impact of the real
exchange rate on FDI.
2 The Data and the MZIP Model
2.1 The FDI data
We examine the average rate of FDI in the US wholesale industry using a panel-data model that
incorporates the interdependence of the states of an industry over time. We use as our basic
empirical framework Campa’s (1993) empirical implementation of the theoretical model
developed by Dixit (1989). Our data sources and specification of empirical variables are also
based largely on Campa although there are some differences which we explain below. Above all,
we use the MZIP model whereas Campa uses the Tobit model. Following Campa, we eliminate
the influence of input origin, production location and output destination on the relationship
between FDI and exchange rate by considering FDI into the United States wholesale trade sector
6
rather than the manufacturing sector. FDI data in the wholesale trade sector is from the United
States International Trade Administration (ITA) publication entitled, “Foreign Direct Investment
in the United States: Transactions.” The ITA publication includes information on the type of
investment, the name and nationality of the foreign investor, the name of the US affiliate, the US
affiliate’s four-digit Standard Industrial Classification (SIC) code and the value of investment in
US dollars.10 However, the ITA publication has many missing observations on the values of
investments due to confidentiality agreements with foreign investors. Because of this, we use the
number rather than the value of FDI in four-digit SIC industries in the wholesale trade sector.11
Following Tomlin, we extend the sample period to cover 1982 to 1994. 12 Due to the
reclassification of some four-digit SIC industry codes after 1987, we have 59 and 69 industries
for 1982-1987 and 1988-1994, respectively. It is important to emphasize that we handle the post-
1987 reclassification by constructing an unbalanced panel data set which contains more SIC four-
digit industries for 1988-1994 than 1982-1987.13 Fourteen additional SIC industries were created
after 1987 while four SIC industries were discontinued after 1987. For each year and each
industry, we enter as our observation the number of FDI. We have 389 non-zero entries or
observations from 1982 to 1994, which show foreign investors from 32 countries making 1,111
investments in the US wholesale trade sector. However, there are years when an industry does not
have FDI recorded in the ITA publication. When there is no FDI in a certain year, we enter zero
as our observation for that year. We have 405 zero observations and they make up 51% of our
total observations. Our sample has a size of 794 observations.
2.2. The MZIP Model for Panel Data
To formally describe the interdependence of FDI over time and to handle the excess zeros in
our data, we adopt a count data model known as the Markov zero-inflated Poisson (MZIP) model
developed by Wang (2001). The MZIP is based on the zero-inflated Poisson (ZIP) regression
models. The ZIP model is used to handle count data with excess zeros but it is not valid when
7
there is interdependence of observations over time. As the ZIP model is a special case of the
MZIP model, we can compare the two models and test for the interdependence of FDI over time
by conducting the likelihood ratio test.
We extend the MZIP model developed by Wang (2001) to panel data with k subjects or
industries. Let }....,,1);,,{( iijijij njtxy = be a sequence of observed data for industry i (i = 1,
….., k), where yij is an observed FDI count associated with time exposure of tij during the jth
period and a vector of covariates ),( )2()1(ijijij xxx = for 2≥j and )2(
1)1(
1 ii xx = where the dimensions
of vectors )1(ijx and )2(
ijx are d1 and d2 respectively. The MZIP model for panel data assumes that:
(i) for an observed FDI count yij for industry i during period j, there corresponds a partially
observed binary random variable, Sij, representing the state of a two-state discrete time
Markov chain with Sij = 1 when yij > 0 and Sij = 0 when yij = 0. Furthermore, we define
the state represented by Sij = 0 as the zero state in which industry i is not favorable to
FDI, and the state represented by Sij = 1 as the Poisson state in which industry i is
favorable to FDI;
(ii) the partially observed binary random vector ).....,,,( 21 jinii SSS for industry i follows the
two-state discrete time Markov chain with transition probabilities defined by
),(log)exp(1
)exp(
)()00Pr(
)1(10)1('
0
)1('0
00)1(
ijij
ij
jiij
xitx
x
ijpSS
ββ
β≡
+=
=== −
(1)
)(1)()01Pr( 0001)1( ijpijpSS jiij −==== − (2)
)x(itlog
)xexp(1)x(exp(
)ij(p)0S0SPr(
)1(ij
'1)1(
ij'
1
)1(ij
'1
11)1j(iij
ββ
β≡
+=
=== −
(3)
)(1)()10Pr( 1110)1( ijpijpSS jiij −==== − (4)
8
where ).....,,(10010 dβββ = and ).....,,(
11111 dβββ = are two unknown parameter vectors
related to the transition probabilities )(00 ijp and )(11 ijp respectively; and
(iii) conditional on Sij = 1, observed FDI count yij follows a Poisson distribution
]),([exp]),([!
1)1,,,( )2()2()2(1 ijij
yijij
ijijijiji txtx
yStxyf ij αλαλα −== (5)
where ),'(exp),(.....,,1,0 )2()2(ijijij xxy ααλ == and ).....,,(
21 dααα = is an unknown
parameter vector; conditional on ,0,0 ≡= ijij yS i.e.,
⎪⎩
⎪⎨⎧
>
===
00
01)0(0
ij
ijijij yif
yifSyf (6)
Under the above assumptions, the likelihood function of the model is
)}1,,,(
)}()([)0()]()({[
)]1,,,()0([
)2(1
11010102
00
11)2(
111)1(
11101
)1(0
=
++=+
=+==
∏
∏
=
=
ijiijij
ijij
n
j
iiiii
ii
k
i
i
Sjtxyf
ijpijpSyfijpijp
StxyfpSyfpl
i
α
α
(7)
Note that while )0Pr( 1)1(
0 == ii Sp and )1Pr( 1
)1(1 == i
i Sp are the unknown probabilities of the
initial states of the Markov chain for industry i, we assume that both initial states are equally
likely and set 5.0)1(1
)1(0 == ii pp . Our Monte Carlo study, which we report below, indicates that
the values of probabilities have little effect on parameter estimates for a large sample. Also, as in
Wang (2001), a sequence of repeated observations over time for a subject is modeled by the
MZIP model for a time series, and the serial dependence of repeated observations for a subject is
described by the hidden Markov chain. The series of repeated observations for different subjects
in a panel data set are assumed to be independent of each other.
As in Wang (2001), we obtain the maximum likelihood estimates using a combination of the
EM algorithm and the quasi-Newton algorithm. Although a numerical method like the quasi-
9
Newton algorithm allows us to directly maximize the likelihood function (equation 7), the EM
method has at least two advantages. First, the EM algorithm is less sensitive to the starting values
of the parameters partially because the functions in the M-step of the EM algorithm are less
complex than the likelihood function. Second, the E-step of the EM algorithm produces the
estimated posterior probabilities which we can use to identify an observation as belonging to
either the zero state or the Poisson state. Finally, we test for the serial dependence of repeated
observations using the likelihood ratio test. We fit the data with an unconstrained model and an
identical model with constraints on the transition probabilities. The second model is nested in the
first, which is why we use a likelihood ratio test for inference.14
2.3 The Monte Carlo Simulation
We use Monte Carlo simulation with 2000 replicates for the MZIP to investigate the variability
of estimates, the effect of the probabilities of initial studies on parameter estimates, and some
finite-sample properties of the maximum likelihood estimates. The number of subjects or
industries k in the sample for each of the four simulations is 10, 20, 40 and 80, respectively, with
each subject or industry having 10 repeated observations. Hence the sample size n = 10k is
accordingly 100, 200, 400 and 800.
Data are generated from the MZIP model which has the Poisson rate function
)5.45.0exp(),( )2(ijij xx +−=αλ for ki .....,,1= and ,10.....,,1=j and the transition probabilities
)5.15.1(log)(00 ijxitijp −= and )22(log)(11 ijxitijp +−= for ki .....,,1= and .10.....,,2=j
Throughout, the covariate xij takes on n uniformly spaced values between 0 and 1, and the first 10
uniformly spaced values are the covariates for subject 1, the next 10 values for subject 2, and so
on. The probabilities of both initial states are 0.5. Note that the time exposure is chosen as
1=ijt for each observation. With these values, the Poisson rate function ),( )2( αλ ijx ranges from
0.38 to 7.39, with a median of 1.67. Both )(00 ijp and )(11 ijp range from 0.27 to 0.73, with a
10
median of 0.5. The parameter values are chosen so that the corresponding functions can be clearly
separated.
We fit each of the 2,000 replicates to the MZIP model using two different probabilities of the
initial states: (1) ;5.0)1(1
)1(0 == ii pp and (2) )1(
0ip is generated by a (0,1) - uniform distribution.
Table 1 reports the simulation results for both probabilities. Clearly, the two probabilities yield
almost identical parameter estimates for all n, suggesting that the probabilities of the initial states
have little impact on parameter estimates when the sample size is at least 100. For all n, both the
mean and median (over 2,000 replicates) of each parameter estimate is close to the corresponding
true value. For n = 100 and 200, the standard errors (over 2,000 replicates) of the parameter
estimates in the transition probabilities are relatively large, suggesting less accuracy in the
estimation of those parameters. This is not surprising since even the standard logistic regression
maximum likelihood estimates have infinite bias in finite samples. The logistic regressions for the
transition probabilities are, of course, harder to fit than the standard logistic regressions because
the state binary random variables are only partially observable. What is encouraging, however, is
that when n > 200 all the parameters can be estimated accurately. This confirms that our
estimation algorithm is reliable. Moreover, we can accurately estimate the parameters in the
Poisson rate function even if we cannot do so for the parameters in the transition probabilities due
to small sample size.
Table 2 reports 95% confidence intervals based on the bootstrap method and normal theory for
four different sample sizes. The lower and upper limits of the bootstrap intervals are the 2.5th
percentile and 97.5th percentile of 2,000 replicates; and the lower and upper limits of the normal
intervals are calculated by the formula: mean ± 1.959964 ×standard deviation. Observe that for all
n the discrepancies between these two types of intervals for parameters in the Poisson rate
function are rather small, suggesting that the maximum likelihood estimates should be normal;
and they are relatively smaller than those for parameters in the transition probabilities. Observe
11
also that the discrepancies between these two types of intervals for each parameter in the
transition probabilities are small for n = 800 and n = 400; the discrepancies for some parameters
in the transition probabilities are slightly significant for n = 200 and n = 100. These results
suggest that the maximum likelihood estimates of the parameters in the transition probabilities are
approximately normal when the sample size is at least 400, and slightly skewed when the sample
size is less than 400. They also suggest that the bootstrap intervals may be more appropriate when
the sample size is not large enough.
2.4. Vectors of covariates
As mentioned earlier, we use as our framework Campa’s empirical model and specify the two
vectors of regressors of the FDI rate function and the transition probabilities for industry i at year
t, )1(itx and )2(
itx , to be the same variables as Campa’s. We initially assume )2(it
)1(it xx = . The
vectors of regressors in our analysis are measures of exchange rate level itR , the rate of change in
the exchange rate itμ , volatility of the exchange rate itσ , sunk costs itk , and variable costs of the
US relative to foreign countries itw . We can summarize Campa’s reduced form function of FDI
projects ity to be estimated, which is instructive for our own MZIP regression, along with the
expected signs of the coefficients, as below.
⎟⎟⎠
⎞⎜⎜⎝
⎛
−−+−+= ititititit
itwkR
y,,,, 2σμ
φ (7)
The definitions and computations of the three exchange rate variables – itR , itμ and itσ – are
based on Campa. More specifically, we define the exchange rate level itR as the average of the
exchange rate in the year of the FDI, itμ as the trend in exchange rate, and itσ as the standard
deviation of the monthly change in the logarithm of the exchange rate. Since itμ and itσ
incorporate firms’ expectations about the future levels of those variables, their computation
12
requires assumptions about how firms form such expectations. As in Campa (1993), we make two
alternative assumptions – perfect foresight and static expectations. The former implies that firms
have perfect forecast expectations of the ex-post value of the exchange rate of the next year. The
latter implies that firms estimate the future exchange rate as the exchange rate of the year
previous to the FDI. Following Campa, the exchange rate variables are computed using monthly
index of foreign currency per US dollar and weighted by the number of FDI.15 Campa provides a
detailed discussion of the FDI weights for the exchange rate variables.16 When the number of FDI
is positive for an industry in a particular year, we calculate an effective exchange rate as the
average of the exchange rate indexes weighted by the number of FDI from a given country.
However, there are two main differences between our and Campa’s computations of the three
exchange rate variables. First, our base year for computing those variables is 1995 whereas
Campa’s base year is 1980. Second, and more importantly, we differ from Campa in terms of the
data source we use to calculate the FDI weights for the three variables when there is no FDI. If
the number of FDI is zero for an industry in a particular year, we calculate an effective exchange
rate using weights based on the total number of firms from a foreign country operating in that
industry from 1973 up to that year. We choose 1973 since it is the first year for which data are
available from the US Department of Commerce, International Trade Administration, “Foreign
Direct Investment in the United States: Completed Transactions”. This data source provides FDI
data for four-digit SIC industries. In contrast, Campa uses a data source which provides three-
digit SIC data, from which he estimates the four-digit SIC data needed to compute the FDI
weights. More specifically, Campa uses the 1980 benchmark survey of the US Department of
Commerce, Bureau of Economic Analysis, “Foreign Direct Investment in the United States:
Operations of US Affiliates: 1977-1980”. Our FDI weights are likely to be more accurate since
our data source provides four-digit SIC data whereas Campa’s data source provides three-digit
SIC data.
13
Let us now look at the variables which are not related to exchange rates, namely sunk costs and
foreign variable costs. While sunk costs are a theoretically important determinant of FDI, they are
difficult to measure empirically. We use the two empirical proxies for industry-specific sunk
costs proposed by Campa. SUNKit is the ratio of fixed assets to net wealth of all US firms in a
four-digit SIC industry and represents all the physical investments that a firm has to incur to
establish itself in the market.17 ADVit is the ratio of media expenditures to company sales by all
US firms in a four-digit SIC industry and represents largely unsalvageable non-physical
investments in advertising, sales force and media promotion.18 We compute both SUNKit and
ADVit exactly as described in Campa. Our measure of the variable production cost or labor cost,
itw , is also the same as Campa’s. However, in computing itw , we use the weighted average of
the unit labor cost indexes of eleven countries with respect to the US rather than ten as in Campa.
Furthermore, we use a more up-to-date version of Campa’s data source, namely the Bureau of
Labor Statistics, News Release 2002, Table 10. The weights are the proportion of FDI from a
given country in each four-digit SIC industry.19
4 Empirical Results
We first examine the interdependence of FDI over time for the case of static expectations,
which means that firms estimate the future exchange rate as the exchange rate of the year
previous to the FDI. To check for evidence of interdependence of FDI over time, we compare the
MZIP and the ZIP regression models using the same covariates for the average FDI rate and for
the transition probabilities and the zero probability in the MZIP and ZIP models respectively. The
parameter estimates of the unrestricted coefficients are reported in Table 3. The results show that
while the parameter estimates of the FDI rate function for both the MZIP and the ZIP models are
quantitatively similar with signs consistent with theoretical predictions, the inference about some
of the parameters differs. For example, at 10% significance level, the coefficient of the unit labor
costs is not significant for the MZIP model, but significant for the ZIP model. This suggests that
14
different models for FDI may lead to different results of the inference about the parameters. The
coefficients of the exchange rate level and trend are positive while the coefficient of the exchange
rate standard deviation is negative. The coefficients of both measures of sunk costs as well as the
labor costs are negative. Since the ZIP model is nested in the MZIP model, we conduct the
likelihood ratio test which gives a test statistic of 72 and a p-value of 0.0000. This rejects the ZIP
model in favor of the MZIP model and shows the interdependence of FDI over time. This also
implies that without consideration of the interdependence of FDI the use of ZIP model may lead
to incorrect inference about the parameters.
For the MZIP with unrestricted coefficients, the t-statistics indicate significance at the 5% level
for the exchange rate trend and 1% level for the exchange rate level. Both measures of sunk costs
are significant at the 1% level. However, the exchange rate standard deviation and unit labor costs
are insignificant even at the 10% level. The bottom half of Table 3 also shows that most of the
regressors of the transition probabilities are insignificant even at the 10% level. Since our results
suggest that the coefficients of the regressors in transition probabilities may be zero, we fit the
data to a restricted MZIP regression with these coefficients equal to zero. The results of the
restricted MZIP model are also shown in Table 3. For comparison, we also run the ZIP
regression with restricted coefficients for the zero probability. As in the unrestricted models, the
inference about some of the coefficients of the regressors differs, and the ZIP model is rejected in
favor of the MZIP model using the likelihood ratio test. We also use the likelihood ratio test to
compare the unrestricted MZIP model with the restricted MZIP model. Since the log-likelihood
ratio test statistic is 16.2 with the p-value of 0.1822, we cannot reject the restricted MZIP in favor
of the unrestricted MZIP. Hence, the restricted MZIP model is more appropriate than the
unrestricted model. For the restricted model, when an industry is favorable to FDI, the average
rate of FDI is given by:
(8) itititit R σμμ 6328.07767.0008.0031.1exp( −++=
)1591.00174.00051.0 ititit ADVSUNKw −−−
15
Using the logit function, we compute the transition probabilities of the restricted MZIP
model 00p , 01p , 11p and 10p to be 0.6966, 0.3034, 0.7161 and 0.2839, respectively.20 This means
that the probability that an industry is in the FDI-unfavorable state in one period when it was in
the unfavorable state in the previous period is thus 69.66% and the probability that an industry is
in the FDI-favorable state in one period when it was in the favorable state in the previous period
is thus 71.61%. Such numbers provide evidence of the interdependence of FDI over time. Our
results also imply that in the long run an industry is in the FDI-unfavorable state 51.66% of the
time, and in the FDI-favorable state 48.35% of the time as the stationary probabilities of the states
of the Markov chain are p0 = 0.5166 and p1 = 0.4835 respectively.21
We now report in Table 4 our results for the case of perfect foresight, which means that firms
have perfect forecast expectations of the ex-post value of the exchange rate of the next year. As
in the case of static expectations, we test for the interdependence of FDI over time by fitting the
data to the ZIP and MZIP models for both the unrestricted and restricted coefficients of the
transition and zero-probabilities. We then conduct the likelihood ratio test to compare the ZIP
and the MZIP models. We get a test statistic of 71.6 for the models with unrestricted coefficients
and a test statistic of 65 for the models with restricted coefficients. For the models with
unrestricted coefficients, we reject the ZIP model in favor of the MZIP model, and hence there is
strong evidence of interdependence in FDI over time.
We report the parameter estimates in Table 4 of the MZIP and ZIP models for perfect foresight.
For the model with unrestricted coefficients, our MZIP regression results for the average rate of
FDI in industries with favorable FDI environments, are largely consistent with theoretical
predictions. The coefficients of the exchange rate level and trend are positive while the
coefficients of both measures of sunk costs as well as the labor costs are negative. The exception,
the positive coefficient for the exchange rate standard deviation, is insignificant, as is unit labor
16
costs. The t-statistics indicate significance at the 1% level for the exchange rate level and 10%
level for the exchange rate trend. Both measures of sunk costs are significant at the 1% level.
Such results are broadly similar to those for static expectations.
Our MZIP regression results for the transition probabilities indicate that none of the regressors
are significant. As noted above, the statistical insignificance of the regressors suggests that we
should restrict their coefficients to be zero, as we did for static expectations. Table 4 shows the
parameter estimates and log-likelihood when we do so.
To compare the two MZIP models reported in Tables 4, we conduct the likelihood ratio test.
Since the test statistic is 6.6 and the p-value is 0.8829, we cannot reject the null hypothesis that
the coefficients of the regressors of the transition probabilities are zero. This confirms that the
restricted MZIP model reported is more appropriate. For industries with favorable investment
environment, all coefficients of the regressors of the rate of FDI other than itσ have the expected
signs and the t-statistic values are also similar although the exchange rate trend is no longer
significant at the 10% level. For the restricted MZIP model with perfect foresight, when an
industry is favorable to FDI, the average rate of FDI is given by:
(9)
Using the logit function, we compute the transition probabilities 00p , 01p , 11p and 10p to be
0.6978, 0.3022, 0.7151 and 0.2839, respectively. The estimated transition probabilities strongly
support the notion that FDI may be interdependent over time. Furthermore, in the long run, each
industry will have a favorable FDI environment with a probability of 51.47% and an unfavorable
FDI environment with a probability of 48.53%.
Our two main empirical findings are the interdependence of FDI over time and a positive
relationship between the exchange rate and rate of FDI inflows for industries which are favorable
to FDI. Our computed Markov transition probabilities suggest that FDI inflows into US
itititit R σμμ 3476.03873.00091.06905.0exp( +++=
)16.00141.00028.0 ititit ADVSUNKw −−−
17
wholesale trade industries may be interdependent over time because of uncertainty over whether
an industry’s environment is favorable or unfavorable to FDI. This uncertainty could be modeled
as a two-state Markov chain. More precisely, if an industry had been favorable to FDI in the
previous period, it is more likely to be favorable to FDI in this period and likewise for the
probability of an industry being unfavorable to FDI.
Our MZIP regression results show that for industries with favorable FDI environments, most of
the coefficients of the regressors of the rate of FDI have the expected signs and some of the
coefficients are highly significant. In particular, under the static expectations and, to a lesser
extent perfect foresight, the exchange rate level and trend have positive and significant impact on
the rate of FDI. This suggests that a stronger US dollar has a positive impact on the rate of FDI
into US wholesale industries. Our finding thus re-confirms the empirical results of Campa for
exchange rate level. Unlike Campa, we find a positive and significant coefficient for the trend in
the exchange rate for the model with static expectation — although it is less significant for the
model with perfect foresight — and a negative but insignificant coefficient for standard deviation
of the exchange rate. Hence, while the coefficients have the expected signs, we do not find
evidence to support Dixit’s (1989) hypothesis that exchange rate uncertainty deters the average
rate of FDI. Our findings also differ from those of Tomlin even for the ZIP regressions. Our ZIP
regression results suggest a positive significant impact of the exchange rate level on the rate of
FDI. This might seem puzzling at first since Tomlin also uses ZIP regressions. However, we
should keep in mind that we use panel data while Tomlin uses pooled cross-sectional data.
Furthermore, we address the issue of post-1987 SIC reclassifications by building up an
unbalanced panel data set and construct the three exchange rate variables on the basis of more
accurate FDI weights. Because of the existence of the interdependence of FDI over time, the
MZIP model is more appropriate than the ZIP model for the analysis of panel FDI data since the
use of the ZIP model may lead to incorrect inference about the parameters.
18
5 Concluding Remarks
Common sense tells us that the real exchange rate has an effect on FDI, just as it has an effect
on international trade. A number of theoretical and empirical studies have examined the
relationship between FDI and the real exchange rate more formally. In particular, Campa
develops an empirically testable model of FDI based on Dixit’s model of investment, which in
turn, is derived from the theory of option pricing in financial economics. Campa’s model predicts,
and the empirical evidence from his Tobit estimation strongly supports, a significant effect of the
real exchange rate on the probability of FDI entry in US wholesale trade industries. However,
using the ZIP model, Tomlin fails to find a meaningful relationship between the exchange rate
and the average rate of FDI.
Our study expands the ZIP model by incorporating the possibility of interdependence of FDI
over time in each industry. To do so, we use the MZIP model, which is based on two-state
Markov chains. For empirical purposes, we extend the MZIP model, which is a time-series
specification, for panel data since we use industry-level panel data for our empirical analysis.
While our data are based largely on Campa, there are some differences. It is also important to
point out that we use an unbalanced panel data set. One of our two main empirical findings is that
FDI is indeed interdependent over time. Such interdependence captures immeasurable and
uncertain factors that affect the state of an industry – whether firms view an industry as favorable
or unfavorable to FDI – and, in turn, these views may be affected by the state of the industry in
the previous period. As mentioned earlier, corporate rivalry may explain such interdependence
since it may induce foreign firms in a particular industry to view investments in an industry as
either favorable or unfavorable in response to competition at home and abroad. Our second main
empirical finding is that when industries are favorable to FDI, the exchange rate-related variables
have positive and mostly significant impact on the rate of FDI inflows. This is especially true for
the level and to a lesser extent the trend. Among the variables not related to the exchange rate,
19
both measures of sunk costs have significant negative effects on FDI.
If FDI is interdependent over time, a model such as the MZIP model that explicitly accounts for
such interdependence is more appropriate for empirical analysis of FDI. Our evidence does
indeed provide strong support for the interdependence of FDI over time, and our study suggests
that the ZIP model may be inappropriate for the analysis of panel FDI data as it may result in
incorrect inference about the parameters. In line with Campa’s findings but in contrast to
Tomlin’s findings, we find that the exchange rate, in particular the level of the exchange rate, has
a significant effect on the rate of FDI inflows into the US. Although there are theoretical grounds
for why the exchange rate level might have either a positive or negative effect on FDI, for the US
wholesale trade sector, our results clearly indicate a positive effects of both the level and the trend
in exchange rate. This implies that a stronger US dollar will promote FDI inflows into the US
wholesale trade sector. At a broader level, our analysis points to a need for future researchers to
incorporate possible interdependence in FDI over time when they examine the determinants of
FDI. Considering such a possibility will strengthen the robustness of their findings.
20
References Baldwin, R., (1989). Hysteresis in import prices: The beachhead effect. American Economic
Review 78, 773-785. Baldwin, R., Krugman, P., (1989). Persistent trade effects of large exchange rate shocks.
Quarterly Journal of Economics 104, 635-654. Blonigen, B., (1997). Firm-specific assets and the link between exchange rates and foreign direct
investment. American Economic Review 87, 447-465. Campa, J., (1993). Entry by foreign firms in the United States under exchange rate uncertainty.
Review of Economics and Statistics 75, 614-622. Caves, R., (1989). Exchange-rate movements and foreign direct investment in the United States.
In: Audretsch, D., Claudon, M. (Eds.), The Internationalization of US Markets. New York University Press: New York.
Caves, R., (1971). International corporations: The industrial economics of foreign investment. Economica 38, 1-27.
Dempster, A., Laird, N., Rubin, D., (1977). Maximum likelihood from incomplete data via the EM Algorithm. Journal of the Royal Statistical Society B 39, 1-38.
Dixit, A., (1989). Entry and exit decisions under uncertainty. Journal of Political Economy 97, 620-638.
Froot, K., Stein, J., (1991). Exchange rates and foreign direct investment: An imperfect capital markets approach. Quarterly Journal of Economics 106, 1191-1217.
Guo, J., Trivedi, P., (2002). Firm-specific assets and the link between exchange rates and Japanese foreign direct investment in the United States: A re-examination. Japanese Economic Review 53, 337-349.
Kiyota, K., Urata, S., (2004). Exchange rate, exchange rate volatility and foreign direct investment. World Economy 27, 1501-1536.
Klein, M., Peek, J., Rosengren, E., (2002). Troubled banks, impaired foreign direct investment: The role of relative access to credit. American Economic Review 92, 664-682.
Klein, M., Rosengren, E., (1994). The real exchange rate and foreign direct investment in the United States: Relative wealth vs. relative wage effects. Journal of International Economics 36, 373-389.
Lambert, D., (1992). Zero-inflated Poisson regression, with an application to defects in manufacturing. Technometrics 34, 1-14.
Nash, J., (1990). Compact Numerical Methods for Computers. Adam Hilger: New York. Pindyck, R., (1991). Irreversibility, uncertainty, and investment. Journal of Economic Literature
29, 1110-1148. Tomlin, K., (2000). The effects of model specification on foreign direct investment models: An
application of count data models. Southern Economic Journal 67, 460-468. Wang, P., (2001). Markov zero-inflated Poisson regression models for a time series of counts
with excess zeros. Journal of Applied Statistics 28, 623-632.
21
Table 1
The Results of the Monte Carlo Simulation Poisson Rate Transition Probabilities α1 α2 β01 β02 β11 β12
True Initial Probabilities Mean
k=80, n=800 -0.6156 4.5720 1.4251 -1.4163 -1.9809 1.9619 k=40, n=400 -0.6192 4.5756 1.4455 -1.4545 -2.0112 1.9879 k=20, n=200 -0.6428 4.6097 1.4458 -1.4566 -2.0820 2.0387 k=10, n=100 -0.6641 4.6378 1.4882 -1.5286 -2.4460 2.3969
Median k=80, n=800 -0.6137 4.5698 1.4244 -1.4168 -1.9527 1.9533 k=40, n=400 -0.6156 4.5711 1.4388 -1.4544 -1.9741 1.9565 k=20, n=200 -0.6291 4.5993 1.4462 -1.4498 -1.9197 1.8877 k=10, n=100 -0.6478 4.6250 1.4542 -1.5037 -2.0156 2.0752
Standard Deviation k=80, n=800 0.0867 0.1057 0.2807 0.4505 0.4664 0.6473 k=40, n=400 0.1265 0.1535 0.3996 0.6414 0.6851 0.9663 k=20, n=200 0.1777 0.2122 0.5651 0.8942 1.0691 1.4536 k=10, n=100 0.2479 0.2965 0.8525 1.3497 2.6273 3.1365
Random Initial Probabilities Mean
k=80, n=800 -0.6143 4.5704 1.4298 -1.4218 -1.9653 1.9435 k=40, n=400 -0.6163 4.5725 1.4496 -1.4583 -1.9978 1.9726 k=20, n=200 -0.6398 4.6072 1.4502 -1.4605 -2.0663 2.0225 k=10, n=100 -0.6623 4.6357 1.4949 -1.5352 -2.4124 2.3591
Median k=80, n=800 -0.6109 4.5649 1.4282 -1.4204 -1.9421 1.9465 k=40, n=400 -0.6114 4.5660 1.4481 -1.4515 -1.9606 1.9470 k=20, n=200 -0.6291 4.5993 1.4462 -1.4498 -1.9197 1.8877 k=10, n=100 -0.6390 4.6162 1.4584 -1.5198 -1.9876 2.0127
Standard Deviation k=80, n=800 0.1177 0.1220 0.2820 0.4515 0.4629 0.6425 k=40, n=400 0.1515 0.1675 0.4000 0.6401 0.6791 0.9583 k=20, n=200 0.2016 0.2281 0.5683 0.8980 1.0680 1.4477 k=10, n=100 0.2694 0.3119 0.8544 1.3478 2.6326 3.1373
22
Table 2 95% Two-Sided Confidence Intervals Based on 2,000 Replicates
Poisson Rate Transition Probabilities Α1 α2 β01 β02 β11 β12
Bootstrap Interval k=80, n=800
lower -0.7937 4.3682 0.8865 -2.2987 -2.9855 0.7351 upper -0.4520 4.7892 1.9882 -0.5358 -1.1214 3.2954
k=40, n=400 lower -0.8685 4.2991 0.6752 -2.7130 -3.4923 0.1979 upper -0.3818 4.8724 2.2225 -0.2218 -0.8790 4.0139
k=20, n=200 lower -1.0055 4.1921 0.3565 -3.2588 -4.5045 -0.5070 upper -0.3046 5.0425 2.5610 0.2582 -0.4202 5.2757
k=10, n=100 lower -1.1911 4.0656 -0.1406 -4.2278 -7.4886 -1.7634 upper -0.2253 5.2474 3.2573 1.0266 0.2711 8.1567
Normal Interval k=80, n=800
lower -0.7856 4.3648 0.8749 -2.2994 -2.8951 0.6932 upper -0.4456 4.7792 1.9754 -0.5333 -1.0666 3.2306
k=40, n=400 lower -0.8670 4.2747 0.6622 -2.7116 -3.3541 0.0939 upper -0.3713 4.8764 2.2287 -0.1974 -0.6683 3.8819
k=20, n=200 lower -0.9910 4.1938 0.3383 -3.2093 -4.1774 -0.8102 upper -0.2946 5.0256 2.5533 0.2961 0.0134 4.8877
k=10, n=100 lower -1.1499 4.0568 -0.1828 -4.1741 -7.5954 -3.7506 upper -0.1783 5.2189 3.1591 1.1168 2.7035 8.5445
23
Table 3 Markov Zero-Inflated Poisson (MZIP) and Zero-Inflated Poisson (ZIP) Regression Results for Static Expectations
Unrestricted Coefficients Restricted Coefficients Variable
MZIP ZIP MZIP ZIP
Constant 1.031** 1.076** 1.031** 1.068** Exchange rate level 0.008*** 0.008*** 0.008*** 0.008***
Trend in exchange rate 0.650** 0.742** 0.777** 0.908** Standard deviation in
exchange rate -0.400 -0.435 -0.633 -0.827
Unit Labor Costs -0.005 -0.005* -0.005 -0.006** Sunk Costs -0.018*** -0.018*** -0.017*** -0.019***
Advertising Expenses -0.163*** -0.169*** -0.159*** -0.163***
Transition Probabilities Zero-Probability Transition Probabilities Zero-Probability
00p 11p p 00p 11p p
Constant -0.066 -0.216 0.232 0.831*** 0.925*** -0.298*** Exchange rate level -0.006 0.007 -0.008*
Trend in exchange rate -2.265 3.148* -2.157** Standard deviation of
exchange rate 11.42* -1.657 5.433*
Unit Labor Costs 0.011 0.002 0.002 Sunk Costs 0.002 0.001 0.005
Advertising Expenses 0.036 0.071 -0.057
Log-likelihood -1,381.6 -1,417.6 -1,389.7 -1,425.5 Notes: ***, ** and * denote significance at the 1%, 5% and 10% levels, respectively. All the variables are described in greater detail in Section 2. 00p ( 11p ) refers to the probability that an unfavorable (favorable) FDI environment in the previous period will remain unfavorable (favorable) in the current period in the MZIP model. Zero-probability, p, refers to the probability of an unfavorable FDI environment in the ZIP model.
24
Table 4 Markov Zero-Inflated Poisson (MZIP) and Zero-Inflated Poisson (ZIP) Regression Results for Perfect Foresight
Unrestricted Coefficients Restricted Coefficients Variable
MZIP ZIP MZIP ZIP Constant 0.713** 0.728* 0.690 0.667
Exchange rate level 0.009*** 0.009*** 0.009*** 0.009*** Trend in exchange rate 0.399* 0.373 0.387 0.393* Standard deviation in
exchange rate 0.279 0.615 0.352 0.694
Unit Labor Costs -0.003 -0.003 -0.003 -0.003 Sunk Costs -0.017*** -0.018*** -0.018*** -0.019***
Advertising Expenses -0.164*** -0.171*** -0.160*** -0.165***
Transition Probabilities Zero-Probability Transition Probabilities Zero-Probability
00p 11p p 00p 11p p
Constant 0.998 -1.122 1.324 0.837*** 0.920*** -0.287*** Exchange rate level -0.001 0.009 -0.010**
Trend in exchange rate 0.990 0.119 -0.574 Standard deviation of
exchange rate -5.690 0.543 -0.027
Unit Labor Costs 0.004 0.010 -0.004 Sunk Costs 0.001 -0.001 0.005
Advertising Expenses 0.019 0.069 -0.055
Log-likelihood -1,388.4 -1,424.2 -1,391.7 -1,424.2 Notes: ***, ** and * denote significance at the 1%, 5% and 10% levels, respectively. All the variables are described in greater detail in Section 2. 00P ( 11P ) refers to the probability that an unfavorable (favorable) FDI environment in the previous period will remain unfavorable (favorable) in the current period in the MZIP model. Zero-probability, p, refers to the probability of an unfavorable FDI environment in the ZIP model.
25
Endnotes 1 Froot and Stein (1991) point out that in the presence of capital market imperfections which make external finance more costly than internal finance, a real depreciation of the US dollar increases the relative wealth of foreign firms and gives them an advantage in buying US assets. Blonigen (1997) develops a theoretical model and finds empirical support for this viewpoint. Furthermore, Klein and Rosengren (1994) note that a weaker US dollar attracts foreign capital into the US by lowering the relative labor costs of the US. 2 See Baldwin and Krugman (1989) and Baldwin (1989). 3 Pindyck (1991) provides an excellent review of the literature on investment decisions under uncertainty. 4 In addition, Campa’s model predicts a positive effect of the rate of change in the exchange rate on FDI, as well as negative effects of both the variable costs of production and sunk costs. 5 According to the literature on foreign investment, the exchange rate’s effect on the investment decision depends on the country where the good is produced, the national source of the inputs used in its production, and the country where the final good is sold. See, for example, Caves (1989). 6 For a full explanation of the empirical measures of all the variables, please refer to Campa (1993). 7 In the limited empirical literature on the link between exchange rates and FDI, Froot and Stein (1991) and Klein and Rosengren (1994) also find evidence of a significant relationship. 8 Caves points out that the existence of local production facilities can give a foreign firm a competitive edge in marketing its product. For example, local production may enable the firm to better adapt its product to the local market and provide ancillary service of higher quality or lower cost. 9 Other than political actions of governments, Caves (1971) notes that another source of uncertainty is the high costs of information about foreign markets which causes foreign firms to make FDI decisions with incomplete information but incomplete information on foreign markets is difficult to measure. Caves also mentions exchange rate changes as a source of uncertainty. However, as in Campa (1993), exchange rate uncertainty may be represented in regressions by the standard deviation of the change in the log of the exchange rate. 10 The types of investments are acquisition and mergers, equity increase, joint venture, new plant, plant expansion, real estate and other categories. 11 Other than Campa (1993), Blonigen (1997), Tomlin (2000) and Klein et al. (2002) also use the number of FDI instead of the dollar values of FDI from ITA publication. 12 The last year in our expanded sample period is 1994 since ITA stopped publishing firm-level FDI transactions that year. 13 The full list of industries for the two sub-periods is available from the authors upon request. 14 Note that the second model is equivalent to the ZIP model of Lambert (1992). Please refer to Wang (2001) for a more comprehensive discussion. 15 Source: International Monetary Fund, “International Financial Statistics,” CD-ROM -2005. 16 Please refer to page 617 of Campa (1993). 17 Source: For 1981, Robert Morris Associates (1982), “Annual Statement Studies”. For other years, Dun’s & Bradstreet, “Industry Norms and Key Business Ratios,” several issues. 18 Source: US Federal Trade Commission, Bureau of Economics, “Statistical Report: Annual Line of Business Report 1977,” Washington D.C., 1985. 19 When there is no FDI, we compute the weights as we do for the three exchange rate variables. 20 For example, 00p = logit(0.8313) = )1/( 8313.08313.0 ee + = 0.6966, 01p = 1 - 00p = 0.3034. 21 After calculating the transition probabilities, we can calculate the stationary probabilities of the two
states of the Markov chain, 0p and, 1p from ⎟⎟⎠
⎞⎜⎜⎝
⎛=⎟⎟
⎠
⎞⎜⎜⎝
⎛×⎟⎟
⎠
⎞⎜⎜⎝
⎛
1
0
1
0
1110
0100
pp
pp
pppp .