Our balanced panel data set covers tourist arrivals in Ethiopia originating from 40 countries1
from 1998 to 2004 (7 years), a total of 280 observations. This is typically important for
tourism studies as it helps to incorporate the features of both the receiving country (Ethiopia)
and the originating ones. The data source for tourist arrivals is Tourism statistics Bulletin
number 8 of Ministry of Culture and Tourism of the Federal Democratic Republic of Ethiopia
(MCT, 2006).
The main data source for the explanatory variables is the 2007 edition of the World
Development Indicators CD/ROM of the World Bank. The CD/ROM is a source of data for
per capita income of the sending countries, the exchange rate between the currencies of
Ethiopia and origin countries, the ratio of Consumers’ Price Indices (CPIs) of Ethiopia and
the origin countries, the ratio of CPIs of Ethiopia and Kenya, the total population of the
sending countries, and the urbanization rate, and number of internet users in Ethiopia. The
length of road networks in Kilometers of Ethiopia is obtained from an unpublished document
of Ethiopian Roads Authority(2008) and the air distance from the capital cities of the origin
countries to Addis Ababa is taken from the website The data for is summarized in Table 3.1. The table shows that the average mean arrival from
a country in a year is 2,307 where the minimum is registered in 1999 from New Zealand (70
tourists) and the maximum is 28,112 in 2004 from USA. This relatively high number could be
attributed to the large number of Ethio-Americans coming to Ethiopia each year. On
average, the tourists covered in this study are from a high-income category as the average
per capita GDP of the sending countries is USD 12,798.84. However, there is a significant
variation in this variable(a standard deviation of USD 3374.54), with the minimum USD
140.45 (Malawi, 2001) and the maximum USD 39,004.86 (Norway, 2004).
The sending countries are: Australia, Austria, Belgium, Canada, Chad, China, Denmark, Egypt, Finland, France, Germany,
Ghana, Greece, India, Israel, Italy, Japan, Kenya, Korea, Kuwait, Malawi, Mali, Netherlands, New Zealand, Nigeria, Norway,
Pakistan, Philippines, Rwanda, Russia, South Africa, Sudan, Sweden, Switzerland, Tanzania, Turkey, Uganda, UK, USA,
Yemen.
Ghana, Greece, India, Israel, Italy, Japan, Kenya, Korea, Kuwait, Malawi, Mali, Netherlands, New Zealand, Nigeria, Norway,
Pakistan, Philippines, Rwanda, Russia, South Africa, Sudan, Sweden, Switzerland, Tanzania, Turkey, Uganda, UK, USA,
Yemen.
The average distance from Addis Ababa and the capital cities of the countries of origin is
5485.82 kilometers (a bit smaller than the distance between Addis and Paris, 5571.15 kms) Yemen is the nearest country included in the study (922.91 kilometers) and New
Zealand is the farthest (14415.65kilometers).
3.2 Estimation Methodology
For dynamic panel data sets (where the model includes lagged dependent variable), the
lagged dependent variable is by construction correlated to the unobserved country specific
error term causing biases in Ordinary Least Squares(OLS) estimators (Casseli et al, 1996).
For such models, Generalized Methods of Moments (GMM) estimators of Arellano and Bond
(1991) and Blundell and Bond (1998) have great advantage of avoiding endogeneity and
omitted variable bias.
The following illustration of how the systems GMM estimators of Blundell and Bond (1998)
works for a dynamic panel model like ours is based on Levine et al. (2000) and Beck et al.
(2000).
Consider the following equation.
yit = α yit− + βX it +ηi + ε it
∧
1 , where
∧α
= 1+α (1)
where, yit is the logarithm of tourist arrivals; X is the set of explanatory variables (other than
lagged tourist arrivals);η is unobserved country specific effect; ε is the error term; and the
sub-scripts i and t represent country of origin and time period, respectively.2
Casseli et al.(1996), showed the correlation between yit and η makes yit endogenous and
thus OLS estimation of equation(1) results in biased estimates. To avoid such biases, let us
take the first differences of equation (1).
yit − yit−1
=
∧α
( yit−1 - yit −2 ) + β ' ( X it − X it−1) + (ε it − ε it−1) (2)
Applying OLS on equation (2) gives us the fixed effects estimators. However, fixed effects
estimators might be prone to bias for two reasons. First, the explanatory variables in the set
X might be endogenous. Second, in the period t −1, the lagged dependent variable ( yit−1 -
yit −2 ) is correlated with the new error term, (ε it − ε it−1) .
In lieu of the fact that it is usually difficult to find good instrumental variables and these
instrumental variables might be jointly endogenous, Arellano and Bond (1991) suggest the
use of internal instruments, defined as instruments based on lagged values of explanatory
variables. Under the assumptions that the error term ε is not serially correlated, and the
explanatory variables are weakly exogenous (uncorrelated with future realization of the error
term), the GMM dynamic panel estimator by Arellano and Bond (1991) uses the following
moment conditions.
5485.82 kilometers (a bit smaller than the distance between Addis and Paris, 5571.15 kms) Yemen is the nearest country included in the study (922.91 kilometers) and New
Zealand is the farthest (14415.65kilometers).
3.2 Estimation Methodology
For dynamic panel data sets (where the model includes lagged dependent variable), the
lagged dependent variable is by construction correlated to the unobserved country specific
error term causing biases in Ordinary Least Squares(OLS) estimators (Casseli et al, 1996).
For such models, Generalized Methods of Moments (GMM) estimators of Arellano and Bond
(1991) and Blundell and Bond (1998) have great advantage of avoiding endogeneity and
omitted variable bias.
The following illustration of how the systems GMM estimators of Blundell and Bond (1998)
works for a dynamic panel model like ours is based on Levine et al. (2000) and Beck et al.
(2000).
Consider the following equation.
yit = α yit− + βX it +ηi + ε it
∧
1 , where
∧α
= 1+α (1)
where, yit is the logarithm of tourist arrivals; X is the set of explanatory variables (other than
lagged tourist arrivals);η is unobserved country specific effect; ε is the error term; and the
sub-scripts i and t represent country of origin and time period, respectively.2
Casseli et al.(1996), showed the correlation between yit and η makes yit endogenous and
thus OLS estimation of equation(1) results in biased estimates. To avoid such biases, let us
take the first differences of equation (1).
yit − yit−1
=
∧α
( yit−1 - yit −2 ) + β ' ( X it − X it−1) + (ε it − ε it−1) (2)
Applying OLS on equation (2) gives us the fixed effects estimators. However, fixed effects
estimators might be prone to bias for two reasons. First, the explanatory variables in the set
X might be endogenous. Second, in the period t −1, the lagged dependent variable ( yit−1 -
yit −2 ) is correlated with the new error term, (ε it − ε it−1) .
In lieu of the fact that it is usually difficult to find good instrumental variables and these
instrumental variables might be jointly endogenous, Arellano and Bond (1991) suggest the
use of internal instruments, defined as instruments based on lagged values of explanatory
variables. Under the assumptions that the error term ε is not serially correlated, and the
explanatory variables are weakly exogenous (uncorrelated with future realization of the error
term), the GMM dynamic panel estimator by Arellano and Bond (1991) uses the following
moment conditions.
No comments:
Post a Comment