The role of randomization tests in vegetation boundary detection with moving split-window analysis.

Aim Moving split-window ( MSW) analysis is a frequently used tool for vegetation boundary detection. The statistical test of the method is, however, not satisfactory. We aimed to identify reliable confidence limits of MSW statistics. Methods Multivariate transect data (representing forest, grassland and forest-steppe habitats and some artificial data sets) were analysed using five dissimilarity functions and two randomization methods to find the most efficient procedure. We tested the normality of distribution of the dissimilarity values, and then compared the effect of plot randomization and random shift randomization on the power of the statistic. We evaluated the scale dependence of the dissimilarity/distance values and confidence limits. Results The distribution of expected dissimilarity values deviated from a normal distribution, and in all analyses appeared to be skewed to the right. The rate of deviation depends on the data set, on the spatial scale (window size) and on the type of randomization. Expected mean dissimilarity decreased considerably with increasing window width when plot randomization was applied, but was more balanced if random shift was used. Consequently, normalized dissimilarity values increased rapidly with window width in plot randomization, causing extreme scale dependence and false significance. The analysis with random shift distinguished between significant and non-significant peaks at every window size. Conclusions We suggest the use of the random shift permutation as a new method in the MSW analysis, since it minimizes the scale dependence of standardized dissimilarity values, making a clear distinction between significant and non-significant discontinuities. [ABSTRACT FROM AUTHOR]