HIERARCHICAL STRUCTURE OF GENETIC DISTANCES - EFFECTS OF MATRIX SIZE,SPATIAL-DISTRIBUTION AND CORRELATION STRUCTURE AMONG GENE-FREQUENCIES

Geographic structure of genetic distances among local populations with in species, based on allozyme data, has usually been evaluated by esti mating genetic distances clustered with hierarchical algorithms, such as the unweighted pair-group method by arithmetic averages (UPGMA). Th e distortion produced in the clustering process is estimated by the co phenetic correlation coefficient. This hierarchical approach, however, can fail to produce an accurate representation of genetic distances a mong populations in a low dimensional space, especially when continuou s (clinal) or reticulate patterns of variation exist. In the present s tudy, we analyzed 50 genetic distance matrices from the literature, fo r animal taxa ranging from Platyhelminthes to Mammalia, in order to de termine in which situations the UPGMA is useful to understand patterns of genetic variation among populations. The cophenetic correlation co efficients, derived from UPGMA based on three types of genetic distanc e coefficients, were correlated with other parameters of each matrix, including number of populations, loci, alleles, maximum geographic dis tance among populations, relative magnitude of the first eigenvalue of covariance matrix among alleles and logarithm of body size. Most coph enetic correlations were higher than 0.80, and the highest values appe ared for Nei's and Rogers' genetic distances. The relationship between cophenetic correlation coefficients and the other parameters analyzed was defined by an ''envelope space'', forming triangles in which high er values of cophenetic correlations are found for higher values in th e parameters, though low values do not necessarily correspond to high cophenetic correlations. We concluded that UPGMA is useful to describe genetic distances based on large distance matrices (both in terms of elevated number of populations or alleles), when dimensionality of the system is low (matrices with large first eigenvalues) or when local p opulations are separated by large geographical distances.