Ibero-Amerika Institut für Wirtschaftsforschung

Instituto Ibero-Americano de Investigaciones Económicas

Ibero-America Institute for Economic Research
(IAI)
Georg-August-Universität Göttingen
(founded in 1737)

Diskussionsbeiträge · Documentos de Trabajo · Discussion Papers

Nr. 140

Problems in Applying Dynamic Panel Data Models:

Theoretical and Empirical Findings

Felicitas Nowak-Lehmann D., Dierk Herzer,

Sebastian Vollmer, Inmaculada Martínez-Zarzoso

May 2006

Platz der Göttinger Sieben 3 ⋅ 37073 Goettingen ⋅ Germany ⋅ Phone: +49-(0)551-398172 ⋅ Fax: +49-(0)551-398173
e-mail: [email protected] ⋅ http://www.iai.wiwi.uni-goettingen.de
 1

Problems in Applying Dynamic Panel Data Models:

Theoretical and Empirical Findings

By

Felicitas Nowak-Lehmann D.: Ibero-America Institute for Economic Research, fnowak@uni-

goettingen.de,
Dierk Herzer : Ibero-America Institute for Economic Research, [email protected]
Sebastian Vollmer: Ibero-America Institute for Economic Research,
[email protected]
Inmaculada Martínez-Zarzoso: Department of Economics, Applied Economics,
[email protected]
 2

Abstract
The objective of this paper is twofold: First, the applicability of a widely used dynamic
model, the autoregressive distributed lag model (ARDL), is scrutinized in a panel data setting.
Second, Chile’s development of market shares in the EU market in the period of 1988 to 2002
is then analyzed in this dynamic framework, testing for the impact of price competitiveness on
market shares and searching for estimation methods that allow dealing with the problem of
inter-temporal and cross-section correlation of the disturbances. To estimate the coefficients
of the ARDL model, FGLS is utilized within the Three Stage Feasible Generalized Least
Squares (3SFGLS) and the system Generalized Method of Moments (system GMM) methods.
A computation of errors is added to highlight the susceptibility of the model to problems
related to underlying model assumptions.

Keywords:
dynamic panel data model, autoregressive distributed lag model; pooled 3Stage Feasible
Generalized Least Squares estimation, panel GMM estimation, market shares
JEL: F14, F17, C23
 3

Problems in Applying Dynamic Panel Data Models:

Theoretical and Empirical Findings

1. Introduction
In this paper we utilize an autoregressive distributed lag model (ARDL) to estimate the
dynamics of Chile’s market shares in the EU market. This dynamic model has been adapted
from studies of inter alia Balestra and Nerlove (1966), Baltagi and Levin (1986), Arellano and
Bond (1991), Blundell et al. (1992), Islam (1995), Ziliak (1997). Cable (1997) applied an
ARDL to market share behavior and mobility in the UK daily newspaper market. A common
feature of all these studies (and many more studies of this kind) is that the dynamic
relationship between dependent and independent variables is captured by a lagged dependent
variable thus leading to an autoregressive distributed lag models. This is “the” standard
dynamic model that is applied to panel data, as described in Baltagi (2005).
The main aim of this paper is to examine the applicability of the ARDL from a theoretical and
an empirical point of view. From a theoretical point of view, the structure and origin of this
widely used autoregressive distributed lag model are analyzed. From am empirical point of
view estimation problems of the ARDL are illustrated by an empirical application to Chile’s
market shares in the EU market. We distinguish three types of caveats. The first caveat is
related to the theory and refers to the underlying assumptions of the ARDL and the underlying
geometric lag structure. The second caveat deals with the time series properties of the series
and the autocorrelation problem present in most panel data sets. Finally, the third caveat
centers around the endogenity of the lagged dependent variable on the right hand side and the
endogenity of standard instrumental variables in the presence of serial autocorrelation.
The first type of problems arises because the ARDL is derived from a geometric lag (Koyck
lag) model which presumes that all right hand side variables impact on the dependent variable
in exactly this geometric form (Koyck, 1954). The reason for transforming the geometric lag
model into an ARDL is that the geometric lag model is non-linear in its parameters. Non-
linearity in the parameters was considered problematic for estimation in former times.
Nowadays, modern computer software allows one to apply non-linear least squares to the
Koyck-lag model so that this transformation could be regarded as superfluous. Nonetheless,
ARDL continues to be “the” preferred dynamic model since it is so appealing to summarize
the impact of all regressors (lagged and unlagged) in just one variable, namely the lagged
dependent variable! However, derivation of the ARDL from the geometric lag model clarifies
how restrictive the autoregressive ARDL could be.
 4

The second type of problems is basically due to non-stationarity of the data entering the panel
analysis. Non-stationarity leads to serial correlation, a problem that has to be dealt with if
present. Panel unit root test and panel autocorrelation test must therefore be applied before
running regressions to check for the presence of autocorrelated disturbances.
The third type of problems arises only when problem 2 applies. In the presence of
autocorrelated error terms additional estimation problems caused by ”derived endogenity”
appear. The lack of exogenity of the lagged dependent variable and/or standard instrumental
variables is the logical consequence of serial correlation. To tackle these estimation problems,
the dynamic panel data model of Chile’s market shares is estimated by both the Three Stage
Least Squares (3SLS) and the Generalized Method of Moments (GMM) in combination with
Feasible Generalized Least Squares (FGLS) to deal with the problem of endogenity and of
autocorrelation of the residuals across cross-sections and over time.
The critical examination of the preconditions, the applicability on panel data and the
problematic nature of ARDL is considered as the main task of the paper and is pursued in
three steps: First, we strive to clarify what it means to have the geometric lag as underlying
lag structure and to outline the conditions under which a transformation from a Koyck-lag
model into an ARDL would be possible. Second, the estimation problems surrounding the
ARDL in the presence of autocorrelated disturbances, taking for granted that the ARDL is the
true model, are discussed and two estimation methods 3SLS and system GMM are proposed.
Third, ARDL is then actually applied to panel data, even though one has to be careful doing
so. This last step is completed with an error analysis.
From an applied economist’s point of view the objective of the paper is to analyze Chile’s
market share in the EU-market on a sectoral level in the period of 1988 to 2002 by applying
panel time-series techniques. The widely used ARDL model is built with six cross-sections
(EU countries) and fifteen annual observations for the seven most important export sectors of
Chile (fish, fruit, wine, ores, wood, pulp of wood and copper). According to this model
market shares are determined by Chile’s and its main competitors’ relative prices in the EU
countries and an unobserved variable, such as strategic behavior. Price competitiveness is
considered a decisive determinant of Chile’s market shares since Chile’s successful export
products are rather homogeneous products (fish, fruit, beverages, ores, copper, and wood and
products thereof).

The paper is set up as follows. In section 2 the derivation of the model and the assumptions of
the ARDL are analyzed and discussed. Section 3 contains some background information on
 5

Chile’s market shares in the EU to motivate the model and its empirical application. Section 4
presents the application of the ARDL to Chilean market share data and an error analysis.
Section 5 concludes.

2. Building an ARDL with Panel Market Share Data: Some Caveats

2.1 How to Model Market Shares? Econometric Model Versus Purely Stochastic Model
Following Sutton (2004), there are two contradicting views on the development of market
shares over time: The first goes back to Alfred Chandler and asserts that market shares are
robust over time and that leadership tends to persist for a ‘long’ time. The second view,
propagated by Schumpeter, emphasizes the transience of leadership positions. Schumpeter
labels those positions temporary monopolies created by invention and innovation. However,
there is no benchmark for long or short leadership positions (2002 Japan Conference, 2005).
We will test the relevance of these hypotheses by means of panel unit root tests. If market
shares turn out to be stationary (I(0)), this will indicate that they are robust and persistent
during the period of 1988 to 2002. However, if they result to be non-stationary, then we will
conclude that the Schumpeter hypothesis cannot be rejected by the 1988-2002 data.
There are also two approaches of modeling market shares: According to the first approach,
market shares are basically purely stochastic, according to the second approach market shares
are influenced by hard economic factors such as prices, marketing expenditure, number and
strength of competitors etc. When modeling market shares, Sutton (2004) chooses an eclectic
approach. Favoring the idea of building a stochastic model, he enriches the model by
industry-specific features (e.g. a strategic representation of firms’ competitive responses to
market share changes). However, it has to be kept in mind that strategic behavior is very often
intrinsically unobservable. In contrast to Sutton, we put less emphasis on the stochastic nature
of market shares and stress the role played by sectoral real effective exchange rates that can
be treated as an industry-specific feature. We believe that exchange rates, cost differentials,
tariffs and subsidies are important ‘hard’ factors explaining market shares over time.
Therefore, we will build a dynamic econometric model in which price competitiveness is
considered as decisive for the competitive position. Since strategic behavior is difficult to
model, we assume that strategic behavior and sector-specific characteristics are incorporated
in the residuals of the regression model.
 6

2.2 The Dynamic Econometric Market Share Model: The ARDL and the Restrictiveness
of its Assumptions

An autoregressive distributed lag model will be utilized as dynamic model. Since this model
serves as standard dynamic model in panel data analysis, its (general) applicability will be
carefully scrutinized. Our objective is to discuss the preconditions for its applicability and its
limitations by deriving this model. The ARDL approach has been applied in a multitude of
cases and to diverse issues, such as the dynamic demand for natural gas, the dynamic demand
for drug-like products, such as cigarettes, a dynamic model of employment, a dynamic model
for growth convergence, a dynamic lifecycle labor supply model or a dynamic gravity model
(see Balestra and Nerlove (1966), Baltagi and Levin (1986), Arellano and Bond (1991),
Blundell et al. (1992), Islam (1995), Ziliak (1997), Kim et al. (2003)). Finally, it has also been
applied to market share behavior by Cables (1997).

Cable (1997) proposes to model market shares using an autoregressive distributive lag model
(ARDL)1. He selects a first order autoregressive model with a 1-period lagged endogenous
variable2, in which prices and advertising share are the explanatory variables for UK’s
national daily newspapers.

We utilize and modify this model in the following way: Chile’s market share in a specific
sector is determined by Chile’s price advantage (in terms of EU-Chilean producer prices and
EU protection) and Chile’s competitors price advantage in the EU market. In this model,
changes in the real effective exchange rate in the more distant past have a smaller impact on
changes in market shares than exchange rate changes of the more recent past. This assumption
can be very plausible, but must be verified by the underlying data. As will be shown this
model originates from a geometric lag model (Equation (1)) and enables one to model the
reaction of market shares in the short, medium and long run. The lag length is expressed by k
in our model.

Chile’s market share in country i in sector s at time t in the geometric lag approach is
modelled using a log-log-specification:

l shwist = α is + β 0 λ0 lreerist + ... + β 0 λk lreerist−k + γ 0 λ0 lreer *ist +... + γ 0 λk lreer *ist −k + µist (1)
where

1
First order autoregressive model.
2
There are two types of autoregressive distributed lag models: the geometric lag model and the transfer function
model, also known as ARMAX model (for a good description see Greene, 2000)
 7

i = 1, 2,…, 6 represents the cross-sections: FRA, NDL, DEU, ITA, GBR and ESP (according
to World Bank abbreviations);
t = 1988, 1989, …, 2002 are years (annual observations)
s = 03, 08, 22, 26, 44, 47 and 74 are the sectors (according to the two digit HS classification)
lshwist stands for Chile’s market share in EU country i in sector s at point t. lreerist is Chile’s

real effective exchange rate, prevailing in country i and in sector s and lreer *ist is Chile’s

competitor (*) real effective exchange rate, prevailing in country i and in sector s.
Market shares in a specific sector (s) are computed as ratio of Chile’s sectoral exports (X in
the numerator) and EU country i’s imports from the world M.i = MEU+Mnon-EU (in the
denominator). Due to unsubstantial trade volumes, we consider only Chile’s market shares in
France (FRA), the Netherlands (NDL), Germany (DEU), Italy (ITA), UK (GBR), and Spain
(ESP). Market shares are computed for seven sectors at the two-digit HS chapters, namely
fish (03), fruit (08), beverages (22), ores (26), wood (44), pulp of wood (47) and copper (74).
Sources of the data and generation of the data are described in Appendix 1. The period
covered goes from 1988 to 2002. Thus, we obtain a maximum of 6 cross-sections and 15
years, resulting in a maximum of 90 observations per sector. The number of observations
varies depending on the sector studied.

As to the coefficients and the disturbance in this type of model it is assumed that: 0 p λ p 1
and that λ is the same for all regressors. Having the same λ for all the regressors we can
transform eq. (1) into an autoregressive distributed lag model, otherwise this will not be
possible. Besides, if λ is the same, lag length k must be the same for all regressors, too (see
Figure 1).

It is furthermore assumed that β i = β 0 λi , γ i = γ 0 λ i and µ ist ~N(0; σ µ2 ).

A model that follows the above-mentioned restrictions can be transformed into the so-called
first order autoregressive model which is characterized by a lagged endogenous variable on
the right hand side (see Kelejian and Oates, 1981; Greene, 2000 and Nowak-Lehmann D.,
2004).
By lagging eq. (1) by 1 period, multiplying through with λ we obtain

λlshwist −1 = α is λ + β 0 λ1lreerist −1 + ... + β 0 λk +1lreerist −k −1 + γ 0 λ1lreer *ist −1 +...

(1’)
+ γ 0 λk +1lreer *ist −k −1 + µist −1λ
 8

By substracting (1’) from (1) and by suppressing β 0 λk +1lreerist −k −1 and

γ 0 λk +1lreer *ist −k −1 since both terms become very, very small with large k, we obtain an
autoregressive distributed lag model (eq. (2)) which is very similar to the partial adjustment
model3 (Kim et al., 2003):

lshwist = α * is + β 0 is lreerist + γ 0 is lreer*ist + λis lshwist-1 + vist (2)

with α * is = α is (1 − λ ) and vist = µ ist - λ µ ist −1 following a normal distribution N(0; σ v2 ).

However, if λ becomes relatively large (say λ = 0.9) and if the lag length k is short (say k =
2), suppression of the above-mentioned terms turns out to be very problematic since about 70
% (i.e. 0.93) of the impact of the lagged variables would be neglected. This will be shown in
detail in section 4.1 and 4.2 in tables 4 and 6 which contain the error analysis.
A short lag length might constitute a problem when working with annual data, but might be of
minor importance when working with monthly or daily data where the lag length is usually
larger.
Besides and most important, eq. (2) is very restrictive, since the underlying model is a
geometric lag model (eq. (1)) which is known to be of the following form.

Figure 1:

The geometric lag distribution for a parameter bi

Model (2) assumes not only a geometric reaction of the market share (lshw) with respect to

relative prices ( β i and γ i must follow a geometric lag) in all six importing countries i under

3
The partial adjustment model would look like eq. (2*):
 9

investigation, but it assumes exactly the same (as measured by λi ) geometric reaction of lshw

with respect to changes of all the regressors (both lreer and lreer*). In our case, as well as in
many other studies using the ARDL, this assumption cannot be justified by the data for all
regressors. Also, this specific geometric reaction does not always apply to all countries under
study. These issues become even more crucial when the number of cross-sections gets bigger
and when we have some more explanatory variables in the model (a model with e.g. 100
countries and 5 regressors).
Moreover, there are many instances in which the assumption of a geometric lag itself will not
be fulfilled. This will be especially the case when reaction lags are present and when therefore
the impact of changes in the current and the preceding periods is smaller than the impact of
changes of earlier periods. In those cases the dynamic model chosen should be a polynomial
lag model which allows one to estimate any lag structure that can be depicted by a polynomial
of order 1, 2,…, p.
Therefore, before applying model (1) or its linear transformation (2) the existence of a
geometric relationship of the coefficients of the independent variables must be scrutinized
very carefully. Incompatibility of the model assumptions with the data will necessarily lead to
inconsistent estimates.
The question that remains unanswered is whether it is more convenient to estimate eq. (1), the
more general geometric lag model, rather than eq. (2), the restricted model. As stated above,
Eq. (1) is non-linear in its parameters, but can be estimated by Non-linear Least Squares
(NLS). By estimating eq. (1) with Non-Linear Least Squares (NLS) together with SUR and
FGLS one will obtain unbiased and efficient estimates if the relative prices (lreer and lreer*)
are exogenous. That is eq. (1) involves no additional estimation problems (beyond the cross-
section and serial correlation) since endogenity of the right hand side variables does not arise
if lreer and lreer* are exogenous. However, Eq. (1) and eq. (2) have in common that the
assumption of a geometric lag must be fulfilled. Non-fulfillment of this assumption will lead
to biased estimates in both models.

2.3 Estimation Techniques for Non-Stationary Panel Data in an ARDL

Assuming for the moment that the underlying assumptions with respect to the geometric lag
of the ARDL model are fulfilled, the time series properties of the series should be checked

lshwist = λ α is + λ β 0 is lreerist + λγ 0is lreer*ist + (1- λis )lshwist-1 + vist; Here it is assumed that the
adjustment to the desired equilibrium level of market share follows a geometric lag.
 10

and a test of autocorrelation of the disturbances should be applied. We proceeded in several

steps:

First, we test the time series properties of the data (all in natural logs). All series, i.e. market
shares (lshw), Chile’s real effective exchange rate (lreer) and Chile’s competitors’ real
effective exchange rates (lreer*) for all country-pairs are subject to tests on non-stationarity
(panel unit root tests) in a first step. This procedure is applied to all seven sectors under
investigation. The possible existence of structural breaks in the series is neglected because
neither fundamental, abrupt changes in economic policy nor tremendous exogenous shocks
could be detected in the period of 1988-2002. The governments of Aylwin, Frei and Lagos
continued the economic policy of the Pinochet government. Consequently, the time series
display no sign of a significant structural shift.

In the statistical analysis we allow for different unit root processes in the panel, i.e. cross-
section specific (country-specific) unit roots. We apply the Im, Pesaran and Shin (2003) panel
unit root test on all series thus considering the possibility of individual unit roots of our panel
data. All variables (lshw, lreer, and lreer*) are non-stationary, integrated of order one (I(1))
with a p-value of 0.00 (exception: lrpcopper with p = 0.02). With respect to market shares,
this finding supports more Schumpeter’s view. According to Schumpeter, gains in market
shares are temporary. Monopolistic positions have to be defended, otherwise they are lost
quite fast. This view seems to especially apply to the fish, fruit, beverages, ores, and the
copper sector. In the wood sectors (44 and 47), market shares appeared more stable (see
figures 5-6 in Appendix 2), but non-stationary according to the tests. Table 1 presents the
results.

Table 1: Results from the Im, Pesaran, Shin (2003) Panel Unit Root Test stating t-bar
values

IPS Panel Unit Root Test Based on Individual Unit Roots

H0: Series has a unit root (series is non-stationary)
Sector 03 Fish and crustaceans, molluscs
Lshw03 Lreer03 Lreer03*=Lreer03nor
Series in levels -1.81 -1.58 -1.94
∆ Series -4.36 -3.42 -3.47
Sector 08 Edible Fruit and nuts
Lshw08 Lreer08 Lreer08*=Lreer08aus
Series in levels -1.68 -1.58 -2.53
 11

∆ Series -5.90 -3.42 -4.11

Sector 22 Beverages, spirits and vinegar
Lshw22 Lreer22 Lreer22*=Lreer08saf
Series in levels -1.62 -1.58 -0.92
∆ Series -4.25 -3.42 -3.34
Sector 26 Ores, slag and ash
Lshw26 Lreer26 Lreer26*=Lreer26bra
Series in levels -1.29 -1.58 -2.26
∆ Series -4.18 -3.42 -7.43
Sector 44 Wood and articles of wood
Lshw44 Lreer44 Lreer44*=Lreer44nor
Series in levels -1.83 -1.58 -1.94
∆ Series -2.80 -3.42 -3.47
Sector 47 Pulp of wood
Lshw47 Lreer47 Lreer47*=Lreer47nor
Series in levels -1.68 -1.58 -1.94
∆ Series -2.93 -3.42 -3.47
Sector 74 Copper and articles of copper
Lshw74 Lrpcopper4
Series in levels -1.34 -1.58 ----------
∆ Series -4.22 -3.42
Note: lshw = market share, lreer = Chile’s real effective exchange rate, lreer* = Chile’s
competitor real effective exchange rate in sectors 03, 08, 22,26,44,47,and 74.

Since all variables are I(1) one could proceed with cointegration analysis and panel
cointegration tests (Pedroni, 1999; Pedroni, 2004). However, cointegration is a long-term
concept, which is not applicable to our short time span. Moreover, with fifteen annual
observations, the power of panel cointegration tests would be too low. But cointegration
analysis is not the only approach that deals with non-stationary series and yields unbiased and
efficient estimates in a dynamic model. FGLS is another possibility as is known from time
series analysis. We exploit the special suitability of FGLS for estimating dynamic models
with panel data (see Stock and Watson, 2003).

4
Lrpcopper serves as an indicator of Chile’s real copper production costs. It is used instead of lreer in the market
share analysis.
 12

FGLS in a panel analysis setting works analogously to the one in the time series setting. The
idea remains the same: Non-stationarity of the series in a regression equation is reflected in
the autocorrelation ρ of the residuals over time. When utilizing annual data first order
autocorrelation is usually at work. This finding found statistical support in our sample.5

The procedure will be described below by abstracting from sectors for a moment. We tested
eq. (3) below, after having computed the residuals νˆ it from the ARDL model (eq. (2))

νˆ it = ∑ kK=1 ρ ikνˆit − k + eit (3),

with eit ~ N(0; σ ei2 ) and k = 1, 2,…K number of lags. Autocorrelation of the residuals is the

mirror image of non-stationary series. The autocorrelation coefficient ρ ik 6 in a way captures

the autoregressive processes (expressed by ρ ik ' , ρ ik ' ' and ρ ik ' ' ' ) prevailing in the series (see
equations (4)-(7)).

In theory we have:

lshwit = ∑ kK=1 ρ 'ik lshw it −k + e’it (4)

lreerit = ∑ kK=1 ρ ' 'ik lreerit − k + e’’it (5)

lreer*it = ∑ K ρ ' ' 'ik lreer*it-k + e’’’it (6)

k =1

∑ kK=1 ρ
iv
lshwit-1 = ik lshw it − k −1 + eivit-1 (7)

Note that FGLS uses a common ρ̂ ik in equations (4)-(7) and transforms the variables
correspondingly.

The FGLS method is applied in three steps: First, eq. (2) is estimated by SUR and the
residuals are computed Second, the order (first order, second order, or p-order) of
autocorrelation ρ̂ ik is estimated applying SUR and significance is tested in eq. (3). 1st order

autocorrelation of the type νˆit = ρ̂ i1 νˆ it −1 turns out to be present and dominant. ρ̂ i1

expresses 1st order autocorrelation. Third, the variables of eq.(1) and (2) are transformed into

lshwzit = lshwit - ρ̂ i lshwit-1,

lreerzit = lreerit- ρ̂ i lreerit-1,

5
ρ is usually well below 1 so that first differencing is a very rough method to get rid of stationarity.
 13

lreerzit* = lreerit*- ρ̂ i lreerit-1*,

lshwzit-1 = lshwit-1- ρ̂ i lshwit-2 and

ε it = νˆit - ρ̂ i νˆit −1

thus generating variables in soft or quasi first differences. Eq. (2) is then estimated on basis of
the transformed variables applying SUR (see Stock and Watson, 2003).

In contrast to the dynamic panel analysis literature (Baltagi, 2005), we stress the time series
properties of the series more than it is usually done. The dynamic panel analysis literature
usually abstracts from autocorrelation of the disturbances in order to elaborate more on the
characteristics of one-way error or two-way error component models in which cross-section
specific and time-specific random effects are present.

We take a different route for several reasons: First, we decide to work with a fixed effects
model since our cross-sections are not randomly drawn, but selected on purpose. Second, we
try to account for time series properties because our time dimension exceeds our cross-section
dimension and therefore time series problems should obtain more weight.

Even though serial correlation in dynamic panel data models is only rarely dealt with in the
econometric literature, the studies by Hujer et. al. (2005), Kim et al. (2003), Sevestre and
Trognon (1996) and Keane and Runkle (1992) dwell on this issue. Keane and Runkle (1992)
and Kim et al. (2003) use the forward filtering 2SLS method (KR estimate), which treats
unknown serial correlation in residual disturbance. This method pretends serial correlation to
be one, which is a very rough estimate. Kim et al. (2003) refine the KR method and work with
the variables in first differences. We, in contrast, estimate the extent of serial correlation in

the sample (our ρ̂ ik )7and then transform the variables correspondingly (in soft or quasi first

differences). Hujer et al. (2005) assume that the residual term follows a moving average
process (eg. MA(1), MA(2)). According to our data however, the residual term follows clearly
an AR(1) process and not an MA(1) process. Panel analyses with macroeconomic data usually
show unit-roots in the series and usually show an autoregressive error process. Therefore,
time series tests on the series and the residuals are a must before starting estimation of the
model.

The AR-error structure has severe consequences on the endogenity of the instruments that can
be used in the 3SLS and the GMM routine. These considerations lead us to an alternative

6
Which is to be estimated since it is unknown.
 14

method of dealing with non-stationary series in a panel regression framework, namely to

FGLS estimation techniques in combination with 3SLS and a GMM with self-selected
instruments. Before interpreting the regression results we will present some facts on Chile’s
market shares for its most important export sectors and emphasize the role of EU and extra-
EU competition. For each sector separate panel ARDLs will be run over the time period of
1988 to 2002, with the EU countries acting as cross-sections in the panel analysis.

3. Economic Background to the Dynamic Model:

3.1 Chile’s Sectoral Market Shares in a Highly Competitive EU Market
Based on 2003 data, the EU is Chile’s first world-wide trading partner. 25% of Chile’s
exports go to the EU and 19% of its imports come from the EU. During the first semester of
2003, mining (predominantly copper) still represented 46% of total Chilean goods exports,
while agriculture, farming, forestry and fishing products represented 13.02%. Trade with
Chile represents 0.45% of total EU trade, placing Chile as 41st in the ranking of EU main
trading partners. Between 1980 and 2002, EU imports from Chile increased from EUR 1.5
billion to EUR 4.8 billion, whilst EU exports to Chile increased from EUR 0.7 billion to EUR
3.1 billion (EU Commission, 2005).
Given the importance of the EU market to the Chilean export industry, Chile was eager to
sign a Free Trade Agreement (FTA) with the EU (3 October 2002) in order to improve its
market access to the EU. From Chile’s point of view, the agreement can be clearly considered
as a means to maintain and/or strengthen its competitive position in the EU market. In the
short run, a reduction or elimination of trade barriers through a FTA and its impact on relative
prices will improve Chile’s competitive position not only with respect to the EU countries but
also with respect to third countries which do not have a FTA with the EU. In the medium to
long run however, the effect of the FTA will be eroded if the EU decides to conclude also
FTAs with e.g. the MERCOSUR’s full members and perhaps some Asian countries.
Given that Chile’s main export commodities comprise copper, fish, fruits, paper and pulp, and
wine and are thus heavily natural resource based, Chile’s actual competitors are already
numerous8: Norway, Russia, Indonesia, Malaysia, the Philippines and Thailand are much like
Chile exporters of timber and rubber. Besides, the South East Asian countries were able to

7
In FGLS the unknown serial correlation coefficient is estimated as described in section 2.
8
Even though Chile can still be considered the most competitive and the least corrupted economy in Latin
America.
 15

strongly increase their light manufactured exports to industrial countries in the last decade.
South Africa, Australia and New Zealand, in the Southern Hemisphere, threaten Chile’s
position as a successful fruit and wine exporter. As far as agricultural products are concerned,
Chile faces stiff competition from the EU countries. UK, Ireland and Norway are Chile’s
main competitors as far as fish exports are concerned. Moreover, China, enjoying low labor
costs, has become a strong exporter of machinery and equipment, textiles and clothing,
footwear, toys and sporting goods and mineral fuels, thus reversing in general terms Latin
America’s competitiveness in textile, clothing and shoe exports. When analyzing the
determination of market shares (section 4, Eq. (2)) we will take account of EU and extra-EU
competition.

3.2 Chile’s Market Shares in the period of 1988 to 2002

In Table 2 we list Chile’s largest export sectors, its export shares and its market shares in the
EU market. In this table the EU market is considered as one market. However, in the
empirical analysis we investigate Chile’s sectoral market shares in specific EU countries.
Table 2: Chile’s seven most important export sectors and their competitive position
HS Sector Annual Export Potential Average
code percentage share extra-EU Market
change of in competitor Share in
exports 20029 10
the EU11
(1988- (1988-
2002) 2002)

03 Fish and 7.2 % 5.2 % Norway 1.22 %

crustaceans,
molluscs
08 Edible fruit 7.5 % 10.0 % Australia, 2.62 %
and nuts South
Africa, New
Zealand
22 Beverages, 44.6 % 7.8 % South 0.77 %
spirits and Africa,

9
Share of Chile’s sectoral exports in total Chilean exports.
10
According to TradeCAN (World Bank, 2002)
11
Share of EU imports from Chile in total EU imports (total EU imports include intra-EU trade).
 16

vinegar Australia
26 Ores, slag 11.9 % 9.1 % Brazil, 3.75 %
and ash Australia,
China
44 Wood and 12.4 % 1.5 % Norway, 0.26 %
articles of Russia,
wood Canada,
Malaysia,
Indonesia
47 Pulp of 13.9 % 6.6 % Norway, 2.89 %
wood Canada,
Russia
74 Copper and 5.4 % 37.0 % South 10.34%
articles Africa,
thereof Canada
Source: EUROSTAT (2003); COMEXT CD ROM, ‘Intra- and Extra-EU Trade, Annual data, Combined
Nomenclature’, European Commission ; own calculations.

All seven sectors experienced remarkable export growth, beverages being the most dynamic
sector. It should be clarified, however, that ‘beverages’ started from a lower level in 1988
than the more traditional sectors such as fruit, wood, pulp of wood, and copper. Copper had
the biggest market share in EU imports with 10.34 %, followed by ores (3.75 %), pulp of
wood (2.89 %), and fruit (2.62 %) in the period of 1988 to 2002.
The development of Chile’s market shares was subject to up and downs in most of the export
sectors. Defending its market shares was no easy business for Chile in the sectors ‘fish’,
‘fruit’, and ‘ores’. As to the sectors ‘beverages’, ‘wood’, ‘pulp of wood’ and ‘copper’ Chile
could maintain or even strengthen its competitive position (see figures 1-7 in Appendix 2).

4. Empirical Analysis of Market Shares within the ARDL

We estimate eq. (2) as a fixed effect model allowing for cross-section specific intercepts ( α i ).
This model can be applied in its unrestricted form (see eq. (2)) by estimating cross-section

specific slope parameters for lreerit, lreerit* and lshwit-1 ( β 0 i , γ 0i and λi ) but given our

limited number of observations in each cross-section we stick to common slope parameters in

all countries. We capture country-specific effects only through cross-section specific
 17

intercepts ( α i ) and try to save degrees of freedom by modeling common slope parameters

( β 0 , γ 0 and λ ) thus estimating eq. (8) for each of the seven sectors:

lshwit = α i + β 0 lreerit + γ 0 lreer*it + λ lshwit-1 + vit (8)

However, as we have seen before, the advantage of having a linear model is at the cost of
having a lagged endogenous variable that is correlated with the disturbance term due to
autocorrelation. When a lagged endogenous variable appears at the right hand side of a
regression equation (as in the geometric lag model of eq. (2) or eq. (8)) and when the
disturbances are autocorrelated (see eq. (3)), the lagged endogenous variable will be
automatically correlated with the disturbance term and thus becomes endogenous. The
endogenity problem of the lagged dependent variable (lshwit-1), which is caused by first order
AR-correlation of the residuals due to non-stationarity of the series, requires either the use of
the Three-Stage Least Squares12 or the use of the GMM (Generalized Method of Moments)
technique. Modern computer programs allow one to generate the variables in soft first
differences directly by adding e.g. an AR(1) term for first order autocorrelation and to
simultaneously apply methods that control for the endogeneity of the regressors.

4.1 Estimating the Impact of Price Competition on Market Shares Utilizing the 3SLS
Approach in the ARDL model

The choice of instruments is crucial for getting consistent estimates in any model, also in the
market share model. We used an indicator of production capacity in real terms as an
instrument for lagged market share (lshwit-1), the difference in PPP-income between Chile and
the importing country as an instrument for lreerit, and the competitor’s real exchange rate in a
transformation that is generally used in polynomial lag models as an instrument for lreer*it. In
Table 3 the impact of price competitiveness on market shares being estimated by Three Stage
Least Squares (3-SLS) is summarized.

12
Three-Stage Least Squares (3SLS) technique is the SUR version of Two-Stage Least Squares (see EViews 5:
User’s Guide, 2004, p. 700)
 18

Table 3: Results for the ARDL market share model estimated by panel-3 SLS

Regression coefficients♣ Goodness of fit measures♦

Equation (2)

Sector- Impact of Impact Adjustm. AR- R2adjusted S.E. of Durbin

results lreer of lreer* Coeff. term regression Watson
stat.
β 03SLS γ 03 SLS λ3 SLS

03 0.82** -0.72 -0.19 0.68*** 0.97 1.02 2.15

short (0.02) (0.19) (0.20) (0.00)
run

08 1.82** -0.14 -0.07 0.69*** 0.99 1.05 1.99

short (0.02) (0.85) (0.70) (0.00)
run

22 -2.09*** 2.01*** 0.62*** -0.08 0.98 1.05 2.04

short (0.01) (0.01) (0.00) (0.64)
run

22 long -6.96*** 6.04*** -------- -------- 0.98 1.05 2.04

run

26 1.83*** 0.06 0.70*** -0.29* 0.96 1.02 2.06

short (0.00) (0.42) (0.00) (0.07)
run

26 long 6.10*** 0.20 --------- -------- 0.96 1.02 2.06

run

44 0.35 -2.35 0.46*** 0.60*** 0.94 1.06 2.36

short (0.76) (0.13) (0.00) (0.00)
run

♣
p-vales in brackets.
♦
In 3SLS the adjusted R2 is negative at times. It is unclear how the goodness of fit measures of the different
cross-sections are to be weighted in order to derive an overall goodness of fit measure. Therefore, the figures
listed should only signal the trend.
 19

44 long 0.65 -4.37 ---------- --------- 0.94 1.06 2.36

run

47 -1.20*** -0.27 0.37*** 0.01 0.99 1.07 1.87

short (0.00) (0.42) (0.00) (0.91)
run

47 long -1.90*** -0.43 --------- -------- 0.99 1.07 1.87

run

74 -0.45*** -------- 0.80*** -0.07 0.99 1.04 2.16

short (0.00) (0.00) (0.66)
run

74 long -2.25*** -------- --------- --------- 0.99 1.04 2.16

run

Under the assumption that the data follow an ARDL model, we find a significant positive
impact of increased Chilean price competition on market shares in the fish (03), the fruit (08)
and the ores (26) sector but no significant negative impact of foreign price competition on
market shares in the seven sectors under study. As to beverages, we find a negative impact of
competitive (low) Chilean prices and a positive impact of low foreign prices on market shares.
Adjustment to the long-run equilibrium was significant in the beverages (22), the ores (26),
the wood (44), the pulp of wood (47) and the copper (74) sector whereas no significant
adjustment took place in the fish (03) and the fruit (08) sector. However, this interpretation of
the results must still be followed with caution, as the error analysis below (Table 4) will show.

Table 4: Error analysis in the 3SLS framework

Sector Computed Error if kmax =1: Error if kmax =2:

adjustment
λ23SLS λ 33SLS
coefficient

λ3 SLS

Fish (03) -0.19 --- ---

 20

Fruit (08) -0.07 --- ---

Beverages (22) 0.62*** 0.38 0.24

Ores (26) 0.70*** 0.49 0.34

Wood (44) 0.46*** 0.21 0.10

Pulp of wood 0.37*** 0.14 0.05

(47)

Copper (74) 0.80*** 0.64 0.51

We can draw several conclusions from the error analysis in Table 4:

(1) The data do not fit the autoregressive lag model in the fish and in the fruit sector. The
λs there carry the wrong sign and are insignificant, but the ARDL requires significant positive
λs that lie in an interval ]0;1[ .

(2) The data can be explained by an ARDL in the rest of the sectors by and large since the λs
lie in an interval ]0;1[ . However, since we work with annual data where the maximum lag
length is usually short (kmax = 2 is very realistic according to the cross-correlations), large
errors will result in the beverages, the ores and the copper sectors where λ is relatively big

and omission of the terms β 0 λk +1 lreer and γ 0 λk +1 lreer* will therefore result in a large
error. For example in the copper sector the error is 64% if kmax is 1 and 51% if kmax is two. I.e.
64% or 51% of the impact of copper prices on the market share in copper are neglected.

(3) Note that the errors are even bigger than computed when we have reasons to assume that
the geometric lag structure does not apply at all instances. Computation of errors in this case
would require that we know the true model.

4.2 Estimating the Impact of Price Competition on Market Shares Utilizing the GMM-
Approach in an ARDL

Alternatively to 3SLS, we estimate the dynamic model by GMM (Holtz-Eakin et al., 1988;
Arellano and Bond, 1991; Caselli, Esquivel, Lefort, 1996; Durlauf et al., 2004). The special
Arellano and Bond (1991) estimator which is based on the model in first differences is not
 21

applicable in our case since the number of instruments created by the GMM technique would
exceed the number of observations. Nonetheless, the classical GMM technique (in levels)
allows one to control for the correlation between the lagged endogenous variable and the
autocorrelated error terms. Judging from the way GMM works, this approach does have a
comparative advantage over 3SLS at controlling endogenity. Control of endogenity is 100%
due to specific model restrictions and therefore a gain in unbiasedness is obtained. However,
efficiency is lost by creating a tremendous amount of moment conditions that have to be
respected. In our case we get 210 moment conditions, i.e. 210 restrictions13, highlighting the
computational burden of this approach (Schmidt et al., 1992).

The classical GMM approach uses lagged variables as instruments for endogenous regressors.
This procedure, however, must be avoided in the presence of autocorrelation of the
distrurbances since it will not eliminate the problem of endogenity under this condition
(Durlauf et al., 2004). Therefore, we do not use lagged variables as instruments of
endogenous regressors, but the instruments of the previous section, such as the difference in
PPP-income between Chile and the importing country, an indicator of production capacity in
real terms and the real exchange rate in a transformation that is generally used in polynomial
lag models.

Table 5: Results for the ARDL market share model estimated by panel-GMM

Regression coefficients♣ Goodness of fit measures

Equation 2

Sector- Impact of Impact Adjustm. AR- R2adjusted S.E. of Durbin

results lreer of lreer* Coeff. term regression Watson
stat.
β 0GMM γ 0GMM λGMM

03 -0.20 -0.78*** 0.64*** -0.24** 0.98 1.04 2.11

short (0.24) (0.00) (0.00) (0.02)
run

03 long -0.55 -2.17*** ---------- ------ 0.98 1.04 2.11

run

13
The number of restrictions is T(T-1) K/2.
♣
p-vales in brackets.
 22

08 2.29* -0.15 -0.15 0.69*** 0.99 1.10 1.98

short (0.07) (0.90) (0.42) (0.00)
run

22 -2.53*** 2.29*** 0.58*** -0.13 0.98 1.06 2.08

short (0.00) (0.00) (0.00) (0.41)
run

22 long -6.02*** 5.45*** -------- -------- 0.98 1.06 2.08

run

26 0.32 -0.17 0.71*** -0.28* 0.89 1.04 2.04

short (0.52) (0.13) (0.00) (0.06)
run

26 long 1.10 0.24 ----------- --------- 0.89 1.04 2.04

run

44 -1.22** -0.98 0.74*** -0.37*** 0.90 1.06 2.26

short (0.04) (0.14) (0.00) (0.00)
run

44 long -4.69** -3.77 ----------- --------- 0.90 1.06 2.26

run

47 -1.07** -0.31 0.40*** -0.05 0.74 0.26 1.87

short (0.05) (0.52) (0.00) (0.80)
run

47 long -1.78** -0.52 ----------- ---------- 0.74 0.26 1.87

run

74 -1.45** -------- 0.37*** 0.49*** 0.99 1.18 2.01

short (0.02) (0.03) (0.00)
run

74 long -2.30 0.99 1.18 2.01

run
 23

Under the assumption that the underlying assumptions of the autoregressive lag model are
fulfilled we can conclude that there is a positive relationship between an increase in Chilean
price competitiveness and market share in the fruit sector (08) and a negative relationship
between low Chilean wine prices (sector 22) and high Chilean copper prices (sector 74) and
respective market shares. Foreign relative prices have a significant impact in the fruit (03) and
beverages (22) sector. In the latter sector the quality aspect in the wine sector is supposed to
be dominant (see Table 5). The role of prices in the wood (44) and the pulp of wood (47)
sector might be severely impeded by illegal logging and illegal imports of wood products.
Illegal logging distorted official trade flows not only of all timber products (roundwood,
sawnwood, veneer, plywood, boards, semi-finished and finished products, and furniture, but
also of pulp, paper, printed products and cellulose)14. This latter statement applies also to the
interpretation of the 3SLS estimation.

Again an error analysis (see Table 6) is made to take account of intolerable inaccuracy when
the actual lag length is short.

Table 6: Error analysis in the GMM framework

Sector Computed Error if kmax=1: Error if kmax=2:

adjustment λ2GMM λ 3GMM

coefficient

λGMM

Fish (03) 0.64*** 0.41 0.26

Fruit (08) -0.15 --- ---

Beverages (22) 0.58*** 0.34 0.20

Ores (26) 0.71*** 0.50 0.36

Wood (44) 0.74*** 0.55 0.40

Pulp of wood 0.40*** 0.16 0.06

(47)

14
Illegal logging is estimated to comprise up to 50% of all logging activity in the key countries of Eastern
Europe and Russia, up to 94% in the key Asian countries, up to 80% in the key African countries and up to 80%
in the key Latin American countries (WWF, 2005; FERN, 2004).
 24

Copper (74) 0.37*** 0.14 0.05

The error analysis of Table 6 reveals three things:

(1) Again, the ARDL does not seem to be the right model to explain market shares in the fruit
sector in a dynamic context. λ is negative and insignificant.

(2) Large errors do occur in the beverages, the ores and the wood sectors given that λ is
relatively large there (see columns 3 and 4).

(3) The errors are even bigger than stated in Table 6 when we have reasons to assume that a
geometric lag structure does not apply at all instances. However, computation of this
additional source of errors is beyond the scope of this paper.

To sum up:

On the one hand, we have the finding that the ARDL-estimations in sections 4.1 and 4.2 have
very respectable adjusted R2 measures and Durbin-Watson (DW) statistics around 2.15

On the other hand, the standard errors of the regressions are relatively high. Moreover, the
error analysis makes clear that the simple dynamic specification in the form of an ARDL
suffers from some drawbacks. The autoregressive lag specification does not seem to apply in
the fish and the fruit sector. Statements in the beverages, the ores, the wood and the copper

sectors are subject to relatively large errors by neglecting the term λk +1 , the impact of
changes in prices and protection16, in the autoregressive transformation.

The estimation results of 3SLS and GMM differ a great deal. This result is puzzling since
exactly the same instrumental variables are utilized in both estimation procedures. However,
3SLS and GMM differ in the number of restrictions that are applied. 3SLS basically works
under the condition to minimize the squared residuals of eq. (2) with IV replacing the right
hand side variables. GMM estimation is built around a multitude of moment conditions (210
conditions) of which some will be relevant and others just irrelevant. A search for relevant
moment conditions does not take place in the GMM routine so that some far off moment
conditions can become binding (see Ziliak, 1997).
 25

5. Conclusions

Assuming that the underlying geometric lag specification can be applied to the data, the
ARDL specification allows one to draw correct inferences about the short, the medium and
the long run. The ARDL specification can be combined with the FGLS technique and is
therefore able to deal with a couple of estimation problems resulting from autocorrelation,
heteroscedasticity and cross-section correlation of the disturbances. Applied to a system of
equations, this technique transforms the variables in the regression equation through
weighting the regressor matrix with a weight matrix that can control for autocorrelation of the
disturbances, for heteroscedasticity of the variance of the residuals and for cross-section
correlation of the disturbances. The endogenity problem is solved with instrumental variables
(IV) in either a 3SLS or a GMM routine. Unlagged IV are utilized to get rid of the endogenity
problem and to obtain unbiased estimates. Furthermore, the 3SLS and the GMM- technique
are able to produce efficient and consistent estimates if ARDL is the true model.

Violation of the geometric lag assumption is to be expected in particular when working with
heterogenous panel data and with multivariate regression models. In this case a polynomial
lag model could be the model of choice if there is not excessive cross-section heterogeneity.
Estimations in the framework of panel error correction models and panel DOLS could be well
advisable, even though these models require much longer time spans to allow for meaningful
panel unit root and panel cointegration tests. Further research is needed on this topic.

15
Even though the DW must be adjusted in the presence of a lagged endogenous, the DW statistic is still able to
roughly indicate problems of autocorrelation and misspecification.
16
All our prices contain sector-specific protection whenever relevant.
 26

References

Ahn, S. and P. Schmidt (1995), Efficient Estimation of Models for Dynamic Panel Data,
Journal of Econometrics 68: 5-27.
Anderson, T.W. and C. Hsiao (1981), Estimation of Dynamic Models with Error
Components, Journal of the American Statistical Association; 589-606
Anderson, T.W. and C. Hsiao (1982); Formulation and Estimation of Dynamic Models Using
Panel Data, Journal of Econometrics 18: 47-82.
Arellano, M. (1989), A Note on the Anderson-Hsiao Estimator for Panel Data, Economic
Letters 31: 337-341.
Arellano, M. and S. Bond (1991), Some Tests of Specification for Panel Data: Monte Carlo
evidence and an Application to Employment Equations, Review of Economic Studies 58:
277-297.
Balestra, P. and M. Nerlove (1966), Pooling cross-section and time-series data in the
estimation of a dynamic model: The demand for natural gas, Econometrica 34: 585-612.
Baltagi, B.H. and D. Levin (1986), Estimating dynamic demand for cigarettes using panel
data: The effects of bootlegging, taxation, and advertising reconsidered, Review of
Economics and Statistics 68:148-155.Baltagi, B. H. (2005), Econometric Analysis of
Panel Data, Third Edition, Chichester […]: John Wiley & Sons, Ltd.
Blundell, R., S. Bond, M. Devereux and F. Schiantarelli (1992), Investment and Tobin’s q:
Evidence from company panel data, Journal of Econometrics 51: 233-257.
Breitung, J. and H. Pesaran (2005), Unit Roots and Cointegration in Panels,
http://www.econ.cam.ac.uk/dae/repec/cam/pdf/cwpe0535.pdf (4 November 2005).
Cable, J. R. (1997), ‘Market Share Behavior and Mobility: An Analysis and Time-series
Application, The Review of Economics and Statistics Vol. LXXIX (1): 136-141.
Durlauf, S., P. Johnson and J. Temple (2004), Growth Econometrics, social systems Research
Institute, SSRI Working Papers 18.
http://www.ssc.wisc.edu/econ/archive/wp2004-18.pdf (06 February 2006).
European Commission (2003), ‘Intra- and extra-EU trade, Annual data, Combined
Nomenclature, Supplement 2’, EUROSTAT, CD ROM of COMEXT trade data base.
EU Commission (2005),’ The EU Relations with Chile’,
http://europa.eu.int/comm/external_relations/chile/intro/ (24 November 2005)
EViews 5: User’s Guide (2004), Quantitative Micro Software, LLC, Irvine, CA.
 27

FERN, Greenpeace, WWF 2004 :

http://www.panda.org/news_facts/newsroom/news.cfm?uNewsID=17214 (28 July 2005).
Greene, W. H. (2000), Econometric Analysis, London: Prentice Hall International (UK)
Limited.
Holtz-Eakin D., W. Newey and H. Rosen (1988), Estimating vector autoregressions with
panel data, Econometrica 56: 1371-1395.
Hujer, R., P. Rodrigues and C. Zeiss (2005), ‘Serial Correlation in Dynamic Panel Data
Models with Weakly Exogenous Regressors and Fixed Effects, Working Paper- March 9,
2005, J.W. Goethe-University, Frankfurt, Germany.
Im, K., M. Pesaran and Y. Shin (2003), Testing for unit roots in heterogeneous panels,
Journal of Economterics 115: 53-74.
Islam, N. (1995), Growth empirics: A panel data approach, Quarterly Journal of Economics
110, 1127-1170.
2002 Japan Conference: A Summary of the Papers, NBER Website, Friday, July 29, 2005;
http://www.nber.org/2002japanconf/sutton.html (29 July 2005).
Judson, R.. and A. Owen (1999), ‘Estimating Dynamic Panel Data Models: A Practical Guide
for Macroeconomists,’ Economic Letters 65: 9-15.
Keane, M. and D. Runkle (1992), ‘On the Estimation of Panel Data Models with Serial
Correlation when Instruments are not Strictly Exogenous,’, Journal of Business and
Economic Statistics 10(1): 1-9.
Kelejian, H.H. and W.E. Oates (1981), Introduction to Econometrics. Principles and
Applications, New York: Harper & Row Publishers.
Kim, M.K., G.D. Cho and W. W. Koo (2003), Determining Bilateral Trade Patterns Using a
Dynamic Gravity Equation, Agribusiness & Applied Economic Report No. 525, Center
for agricultural Policy and Trade Studies. North Dakota State University.
http://www.ag.ndsu.nodak.edu/capts/ documents/NovemberNewsletter2003-2-4-04.pdf (3 March
2006).
Kiviet, J.F. (1995), On Bias, Inconsistency, and Efficiency of Various Estimators in Dynamic
Panel Data Models, Journal of Econometrics 68: 53-78.
Koyck, L.M. (1954), Distributed Lags and Investment Analysis, Amsterdam:North Holland.
Nickell, S. (1981), Biases in Dynamic Models with Fixed Effects, Econometrica 49: 1417-
1426.
Nowak-Lehmann D., F. (2004), Different Approaches of Modeling Reaction Lags: How do
Chilean Manufacturing Exports React to Movements of the Real Exchange Rate?,
Applied Economics 36(14): 1547-1560.
 28

OECD (1997), The Uruguay Round Agreement on Agriculture and Processed Agricultural
Products, OECD Publications, Paris.
Pedroni, P. (1999), Critical Values for Cointegration Tests in Heterogeneous Panels with
Multiple Regressors, Oxford Bulletin of Economics and Statistics 61 (4) Suppl.: 653-670.
Pedroni, P. (2004), Panel cointegration: asymptotic and finite sample properties of pooled
time series tests with an application to the PPP hypothesis, Econometric Theory 20: 597-
625.
Schmidt, P., S. C. Ahn and D. Wyhowski (1992), Comment, Journal of Business and
Economic Statistics 10: 10-14.
Sevestre, P. and A. Trognon (1996), ‘Dynamic Linear Models’, in The Econometrics of Panel
Data. A Handbook of the Theory with Applications, ed. By L. Mátyás and P. Sevestre, pp.
120-144, Dordrecht: Kluwer Academic Publishers, 2nd ed.
Stock, J. (1994), Unit Roots, Structural Breaks, and Trends, Chap. 46 in Handbook of
Econometrics, Vol. IV, edited by R. Engle and D. McFadden, Amsterdam: Elsevier.
Stock, J. H. and M. W. Watson (2003), Introduction to Econometrics, Boston […]: Addison
Wesley.
Sutton, J. (2004), Market share dynamics and the ‘persistence of leadership’ debate. The
Economics of Industry Group/ Suntory and Toyota International Centers for Economics
and Related Disciplines, London School of Economics, 37.
World Bank (2002), TradeCAN (Competitiveness Analysis of Nations) 2002 CD-ROM,
Washington, D.C.
World Bank (2005), World Development Indicators, Data on CD ROM, Washington, D.C.
WTO Trade Policy Review, European Union (1995, 1997, 2000), World Trade Organisation,
Geneva.
WWF (World Wildlife Fund): EU imports of wood based products 2002 (2005):
http://www.panda.org/about_wwf/where we work/europe/problems/illegal_logging/ (29
April 2005)
Ziliak, J. (1997), Efficient estimation with panel data when instruments are pretdetermined:
An empirical comparison of moment-condition estima, Journal of Business and Economic
Statistics 15: 419-431.
 29

Appendix 1
Description of Data
In the following, the variables: sheu, shnoneu, shw, lreer, and lreer* will be described in
original form (not in logs). All data run from 1988 to 2002. Export data (to compute market
shares) were taken from EUROSTAT: Intra- and extra –EU trade, Supplement 2, 2003.
In our case, six cross-sections (6 EU countries: Germany, Spain, France, UK, Italy, the
Netherlands) had basically complete time series.17
(1a) Chile’s market share in the EU with respect to the EU countries: sheu
sheuist measures the share of Chilean exports (x) of sector s in EU country i at time t when
competing against imports (m) from EU countries only:
Sheuist = xist/mEUist
(1b) Chile’s market share in the EU with respect to the non-EU countries: shnoneu
shnoneuist measures the share of Chilean exports of sector s in EU country i at time t when
competing against imports (m) from non-EU countries only:
shnoneuist = xist/mnon-EUist
(1c) Chile’s market share in the EU with respect to the world (EU and non-EU
countries): shw
shwist measures the share of Chilean exports of sector s in EU country i at time t when
competing against imports (m) from EU and non-EU countries:
shwist = xist/mEU+non-EUjst
(2) The Chilean real effective exchange rate: reer
reer is the bilateral real effective exchange rate between Chile and the EU countries (price
quotation system), taking Chile’s point of view. It consists of the real exchange rate (rer) and
basic indicators of EU protection such as EU-tariffs (t) and EU-subsidies (s).
It is computed (all data for ‘rer’ are taken from World Development Indicators CD ROM of
2005) as:
rer = e ⋅ PEU/PChile with

rer = real bilateral exchange rate between Chile and relevant EU country
e = nominal exchange rate (x Chilean Peso/1EUR) between Chile and relevant EU country
PEU = GDP deflator of the EU country under consideration with 1995 as base year (1995 =ˆ
100)

17
Due to missing data, Austria, Belgium, Finland, Luxemburg and Sweden were excluded from the analysis.
 30

PChile = GDP deflator of Chile with 1995 as base year (1995 =ˆ 100)
rer has been adjusted for EU tariff protection (in terms of average EU tariff rate (t)) and non-
tariff protection (in terms of EU subsidy rate (s). Tariff rates prevailing in the EU can be
found in Trade Policy Review European Union, Volume 1, 2000, pp. 88-101 (WTO) and
rough subsidy equivalents are based on qualitative information on non-tariff protection
collected, explained and nicely put together for UNCTAD by Supper (2001).
So we get:
reer = rer ⋅ (1-s)/(1+t)

For the simulations, we assume that the FTA between Chile and the EU brings tariffs down to
zero.
(3) Chile’s competitors (*) real effective exchange rates :reer*
In analogy to (2) the real effective exchange rates of Chile’s main competitors Norway,
Australia, South Africa, Brazil are computed. Nominal exchange rates, Norway’s, Australia’s,
South Africa’s, and Brazil’s GDP deflators are computed from World Development Indicators
CD ROM 2005. Tariff and subsidy rates are borrowed from WTO and UNCTAD (see (2)).
 31

Appendix 2

Figure 1: Chile’s market share in EU’s fish imports with respect to EU and non-EU
competitors in the period of 1988 to 2002
4.0

3.6

3.2

2.8

2.4

2.0

1.6

1.2

0.8
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHEU03 SHNONEU03 SHW03

Figure 2: Chile’s market share in EU’s fruit imports with respect to EU and non-EU
competitors in the period of 1988 to 2002
9

2
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHEU08 SHNONEU08 SHW08

Figure 3: Chile’s market share in EU’s imports of beverages with respect to EU and
non-EU competitors in the period of 1988 to 2002
 32

12

10

0
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHNONEU22 SHW22

Figure 4: Chile’s market share in EU’s imports of ores, slag and ash with respect to EU
and non-EU competitors in the period of 1988 to 2002
8

2
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHW26 SHNONEU26

Figure 5: Chile’s market share in EU’s imports of wood thereof (44) with respect to EU
and non-EU competitors in the period of 1988 to 2002
1.6

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.0
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHEU44 SHNONEU44 SHW44

Figure 6: Chile’s market share in EU’s imports of pulp of wood (47) with respect to non-
EU and world-wide competitors in the period of 1988 to 2002
 33

1
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHNONEU47 SHW47

Figure 7: Chile’s market share in EU’s imports of copper (74) with respect to non-EU
and world-wide competitors in the period of 1988 to 2002

35

30

25

20

15

10

5
88 89 90 91 92 93 94 95 96 97 98 99 00 01 02

SHNONEU74 SHW74