To infer generalizations from samples, • Species that are a long trek from not give a fair estimate of the average seed the rules are to replicate, to randomize a vehicle. mass of the clade. and to circumscribe categories. Suppose Exclusion practices are inevitable we want to compare seedling density on (although some do seem more virtuous Species within a genus upper versus lower slopes of a hill. The than others). Regrettably, they are rarely A very widely used design is to study a set two categories, upper slopes versus lower discussed in publication. If exclusion of species within a genus (or other clade). slopes, need to be defined clearly. Then, practices were listed explicitly more often, The species might be compared across within those categories, quadrats need literature reviewers could benefit greatly. habitats or in relation to selected traits. to be placed at random (each possible The idea is that the basic similarity in location has an equal chance of being Phylogeny and choosing species morphology and in most traits within a sampled) and replicated. Species are often chosen on a phylogenetic genus will make for a better comparison. These inference rules forge the link basis. The three following phylogenetic Working within a genus is seen as between the sampling design of a study designs illustrate some issues that arise: ‘controlling’ for unmeasured traits, at and the scope of generalizations that least partially. Actually though, the can be inferred from it. Boundaries on Phylogenetically independent contrasts within-genus study is not a very good conclusions are set by the decision about across two habitats compromise. On the one hand, it is what categories to compare. Legitimacy For each species chosen in one habitat, unreliable as a method of controlling for depends on the ‘equal-chance-of-being- a related species is chosen in the unmeasured traits, because all kinds of sampled’ principle. The strength of other habitat, such that each pair is species trait can vary within a genus as conclusions can be assessed, because phylogenetically independent (PI) of other well as between genera. On the other replication provides an estimate of pairs within the study. The strength of hand, by not replicating across several variation within categories. this design is to produce a specified clades, the within-genus study fails to give These inference rules are familiar. number of PI contrasts between the two any sense of whether patterns repeat They were thoroughly assimilated, habitats. A serious weakness is that it themselves in multiple divergences. decades ago, into study designs for underestimates the overall difference vegetation sampling and field between the species mixtures in the two In summary experiments. But choosing species to habitats. This is because, in each habitat, To generalize effectively across species, we study is not yet widely seen as a sampling a species is more likely to be included if it should pay more attention to how species problem of the same kind. In functional is closely related to species in the other are chosen for study. There is no single, ecology and ecophysiology for example, habitat. Clades that only occur in one of simple design that works for all purposes, respected papers have compared two or the habitats are less likely to be included but some general recommendations can be three species and interpreted them as in the design. made. First, make an explicit statement representing categories, such as shade about the categories of species being tolerant versus light demanding. Phylogenetically independent contrasts compared, and the exclusion practices My own work has been far from perfect, on a trait used. Second, be clear about the primary so my purpose here is certainly not to These designs choose PI pairs of species aim of the species-selection design. Is the denigrate past work. Rather, I hope to having a minimum difference in some trait aim to obtain maximum divergence on open a discussion about species-sampling (e.g. threefold difference in seed mass). some trait that is a hypothetical causal practices for the future. Generalizations By pre-specifying a wide divergence, variable? Or to maximize the total number across categories of species could become trait-contrast designs give good power of evolutionary divergences for a given much more reliable through improved in relation to the influence of the trait. number of species studied? Or to assess species sampling. A weakness of trait-contrast designs (and the consistency of the effect across The principle for comparing categories also of contrasts across habitat) is that only different clades? Each of these would is clear. All species falling into each the distal PI contrasts can be used, not lead to a different design. Third, try to category are listed, and each species on a divergences deeper in the phylogenetic move at least a little in the direction of list should have an equal chance of being tree. Abstractly, this is because the species the ideal, which is to randomize from the chosen as a replicate. Very few studies have not been chosen at random from the full set of species within each category actually do this. Exclusion practices are phylogeny. More concretely, imagine a to be compared. one reason why. Some types of species that clade in which most species have similar might commonly be excluded are: seed mass but one outlier species has a Mark Westoby • Species with too few individuals at a site seed mass that is fivefold different from Dept of Biological Sciences, Macquarie • Species that decline to germinate on cue the others. This outlier will necessarily be University, Sydney, NSW 2109 Australia • Species with leaves inconveniently chosen as one arm of the PI contrast within e-mail: mwestoby@rna.bio.mq.edu.au small for a gas exchange cuvette the clade. However, the average between • Rare and threatened species this outlier and the other PI partner will Published online: 14 October 2002