Letter to the Editor Statistical Tests and Bootstrapping: Assessing the Reliability of Phylogenies Based on Distance Data1

Difficulties associated with resolving the human-chimpanzee-gorilla trichotomy demonstrate the need to assess the robustness of phylogenetic hypotheses. For distance data, such as those derived from DNA-DNA hybridization and immunological techniques, taking account of measurement error constitutes a special problem when assessing the reliability of best-fit trees. Several different methods, most centered around Sibley and Ahlquist’s hominoid DNA hybridization data (Sibley and Ahlquist 1984, 1987; Sibley et al. 1990), have been proposed for assessing the extent to which experimental error compromises the reliability of phylogenies based on distance data: the Student’s t-test (Sibley and Ahlquist 1984)) jackknifing (Lanyon 1985~)) a delta Qtest [Templeton 1985 (but see Fitch 1986; Ruvolo and Smith 1986; Saitou 1986; Templeton 1986; Felsenstein 1988 )] , the nonparametric Mann-Whitney U test (Fitch 1986 ) , a maximum-likelihood mixed-model analysis of variance ( Felsenstein 1987 ) , and bootstrapping (Krajewski and Dickerman 1990). At present only two of these methods have been used by other than their initiators: jackknifing (e.g., see Lanyon 19856; Sheldon 1987; Bledsoe 1988; Krajewski 1989; Springer and Kirsch 1989; Marshall and Swift, submitted) and the Mann-Whitney U test (e.g., see Caccone and Powell 1989), though Marshall and Swift (submitted) have used a bootstrap technique similar to Krajewski and Dickerman’s ( 1990). Below I demonstrate that the MannWhitney U test (and related tests) may give quite misleading estimates of the robustness of phylogenies derived from distance data. While the data discussed are from DNA hybridization studies, the arguments and conclusions presented are generalizable to all methods that produce distance data. It is concluded that bootstrapping provides the easiest and most powerful approach for assessing the reliabilities of phylogenies derived from distance data, and I present a variation of Krajewski and Dickerman’s ( 1990) method for bootstrapping distance data.