The Impact of Observer Bias on Multivariate Analyses of Vegetation Structure

Abstract

Experimental data were used to evaluate the effects of subjectivity on habitat analyses. Multivariate vegetation observations were made by four observers who independently and repeatedly sampled a series of sites in an oak-maple forest. The data were analyzed univariately using analysis of variance and multivariately using principal component and discriminant function analyses. Observers significantly differed in their measurements on 18 of 20 vegetation variables. Transformation of variables and use of non-parametric methods did not mitigate observer effects whatsoever. Separate principal component analyses (PCA) of each observer's data yielded four sets of axes which weighted variables quite differently. The angular distance between the PCI axes of different observers got as high as 42?. Observer-specific PC ordinations showed that the sizes, shapes and relative positions of site surfaces were all highly variable among observers. Discriminant function analysis (DFA) was shown to be even more observer-sensitive than PCA. This was deduced from fluctuating variable weights and the angles between DFs. Classification success of discriminant models was impaired by observer bias. Some suggestions for field and analytic improvements are presented.