Within the meat sector the structure is even more concentrated. Only Brazil supplies 38%
of the total OECD imported meat in the period between 2001 and 2004. Brazil increased
its import share within the last 10 years by more than 12%. From the other three top
players China and Argentina lost tremendously within this period and only Thailand
managed as well to increase its export share. These four countries alone supplied 83% of
developing countries’ total export to the OECD in the last decade. No other country has
an export share of more than 5%.
In the following cluster analysis the paper goes more into detail with this analysis.
Which countries are winners or losers concerning their export performance? The
analysis begins with the cluster analysis of the fruit/ vegetable sector and continues
with the analysis of the meat sector. In general, it has to be noted that for both sectors,
for the chosen variables (as described in chapter 4) the data does not have a perfectly
clear cluster structure. 11 However, of the available data we regard them to be the best
indicators of export performance. The exploratory nature of cluster analysis possibly
contradicts our assumptions about categories like “winners” and “losers” - their might
be groups which are “similar” in terms of the distance measure in cluster analysis (in
our case the squared Euclidian distance), but are somewhat difficult to interpret, since
they comprise of both slight losers and slight winners. Since the cluster analysis requires
choices of the user at different steps, we put emphasis on distinguishing “losers” and
“winners” as clearly as possible.
The cluster analysis for fruits and vegetables was performed as follows:
Four variables where considered for clustering: 1) the average value of exports (in 2002-
2004), 2) the difference of the export values between the two periods, 3) the ratio
between the two periods, and 4) the coefficient of variation. For the cluster analysis the
total sample of 73 fruit/ vegetable exporters was split in two groups, small and large
exporters. This decision is based on two different reasons. First, the “average” and the
“difference” are highly correlated and therefore not suitable for cluster analysis. This
finding alone indicates to a pattern of more successful, large exporters, or at least a
systematic proportional increase in exports. In addition, the variable “average” is
strongly right screwed with few large exporters and many rather small exporters. As a
consequence the variable "average" was excluded from the cluster analysis and, instead,
the sample was split by the threshold of an average of $ 500,000 thousand according to
the observed distribution. Furthermore, "Thailand" was excluded from the group of large
exporter countries since it was found in the single linkage clustering as an outliner and
treated as an additional cluster.
Cluster analyses were conducted separately for the two samples. To gain an idea about
the potential number of clusters the Ward procedure was used; the final number of
clusters was determined giving higher priority to “difference” and “ratio” than to
“coefficient of variation”. Based on these criteria, a 5- cluster- solution for the group of
59 small exporters, and a 3- cluster- solution for the group of 13 large exporters were
chosen. These solutions were further checked for homogeneity 12. The country grouping
is displayed in Table 1. Clusters are numbered consecutively for each group, starting
from 1 for the small exporters and starting from 10 for the large exporters.
Table 1: Cluster membership - fruit and vegetable exports
11 This became clear from instabilities of solutions using the K-Means algorithm depending on
which of the different clustering variables had a higher contribution to the clustering (this can be
read from the F-value, calculated by ANOVA to estimate how strongly each variable contributes to
the classification). Giving higher priority to a certain variable cannot be forced in a cluster analysis
(unless variables are given different weights), but we considered the F-values in the choice of the
number of clusters in the way that the “difference” and the “ratio” should have a higher
contribution to the classification than “coefficient of variation”.
12 Clusters are „completely homogenous“ according to the criteria that variance within clusters
should be smaller than variance between clusters for every single variable[2] except for cluster 3 of
the “small exporters”, which has a high variance for the variable “coefficient of variation”. Again,
we are less interested in this particular information, and therefore accept the solution.
9