The variation in species composition of trees greater-than-or-equal-to 7.6 cm gbh in thirty-eight plots (mostly c. 0.2 ha in extent) from physiognomically-defined kerangas forest were re-analyzed by principal components analysis ordination (species centering and standardization by sample norm). Analyses were performed separately on basal area abundances, on the densities of trees in three size classes (greater-than-or-equal-to 7.6, greater-than-or-equal-to 30.5 and greater-than-or-equal-to 61.0 cm gbh) and on the density of small and large trees (7.6- < 30.5 and 30.5- < 61.0 cm gbh). A total of 636 taxa were reduced to 381 for analysis, removing those of very low density and plot frequency. Three groups of plots were identified: forest at low elevation, and generally coastal, on deep humus podzols; forest at intermediate elevation on mostly red-yellow podzols with affinities to dipterocarp forest; and forest at high elevation on mostly peaty podzols. The first group was divisible into five subgroups along a drainage gradient, while the more poorly drained plots showing affinities to peat-swamp forest. Forty to eighty of the taxa, depending on the criteria for selection, were sufficient to define a stable, reduced spatial structure of the data matrix. Two subgroups, both coastal on deep podzols, represent the extreme form of kerangas forest per se. A comparison of Agathis borneensis- and Shorea albida-dominated plots revealed few other associated and differentiating taxa. Patterns were clearest from analyses of basal area data and of densities of all and small trees. Ordinations and grouping of plots for small, but not large, tree densities were similar to those for basal area. Different species were differentiated on the basis of the abundance measure, leading to group (tabular) definition of associations in a dual manner. A new system of summarization is presented which combines basal area, density and frequency in a graded hierarchical approach. The association between vegetation and soil type was difficult to unravel because of the limited environmental space sampled. Soil type was confounded with elevation, rainfall and geographical location. A major factor is clay content probably affecting nutrient status and water holding properties. 'Modal analysis' of small tree densities showed clearest patterns in this respect. There were no patterns at the family or genus level, nor in leaf size spectra within kerangas. Problems in the treatment, analysis and summarization of tropical forest data sets are discussed. These problems centre on the scale and intensity of field sampling and the advantages of measuring small trees leading to a dual basal area and density approach. All published studies, including this one, within kerangas forest have used inadequate sampling for the purposes of revealing species changes with respect to soil type and composition.