Classification and Clustering Applications - An Invariant Approach to Statistical Analysis of Shapes - page 259

although more prone to measurement error (Valeri et al.,1998), contain
diagnostic information on the shape of the calvarium.
This clustering analysis has demonstrated two things. First, all four
methods are equally successful in isolating unicoronal synostosis using
cranial base landmarks. This is due to the asymmetric nature of the cra-
nial base in unicoronal synostosis. Second, none of the methods are able
to differentiate symmetric forms of craniosynostosis and deformational
posterior plagiocephaly using landmark data from the cranial base.
6.7 Summary
In this chapter, we have illustrated the use of landmark coordinate
data for the purpose of classification and clustering. There are many
different dissimilarity measures that may be used toward this end; we
have utilized only four. By their very nature, both classification and
clustering involve an arbitrary choice of a dissimilarity measure. There
is no one dissimilarity measure that works best in every conceivable
situation. Our studies indicate that the four different dissimilarity
measures considered in this chapter can behave similarly. The reader
is reminded that in another research context with alternate data, vary-
ing dissimilarity measures can provide dissimilar and even conflicting
results. Our advice is to study the misclassification rate for the partic-
ular research situation using the cross-validation approach described
in this chapter and choose the dissimilariy measure that works best for
it. We also think that data additional to landmark coordinate data
should be incorporated into the calculation of the dissimilarity mea-
sures. This additional information can potentially reduce the
misclassification rate. It might also improve the performance of the
clustering methods.