Discussion 1699 Spatial patterns are often present in results of ecological surveys (Legendre and Troussellier 1988; Legendre and Fortin 1989; Legendre 1993). When surveys span broad geographic areas, sites near one another often have similar levels of species abundance or biomass as a result of similar geologic, climatic, or biotic factors (Legendre and Fortin 1989). This positive autocorrelation at “small” spatial scales means that nearby sites should not be treated as independent replicates in classical statistical analyses (Legendre 1993). Under these conditions, statistical tests are too liberal and prone to type I errors. Despite knowing these pitfalls, many fisheries researchers continue to use classical statistical analyses and ignore the potential influence that autocorrelation could have on interpretations of broad geographic scale patterns (see Hinch et al. 1994) or temporal patterns (Pyper and Peterman 1998) of fish attributes. In a recent issue of the Canadian Journal of Fisheries and Aquatic Sciences, Randall et al. (1995) used data from entire fish taxocenes for 31 lake and 62 river sites distributed throughout the world to suggest that relative to lakes, rivers had considerably greater fish production, density, biomass, and P/B and considerably lower individual fish mass. Although these results were compelling, we are concerned with the robustness of their conclusions. Because many of Randall et al.’s study sites were clustered within specific geographic locales (clusters of streams in Poland (14), England (nine), Malaysia (seven), and New Zealand (three) and clusters of lakes in Canada (12), Scandinavia (seven), and Russia (five)), we suspected that positive small-scale spatial autocorrelation exists in their data. If it does exist, then their statistical approach could be biased towards detecting differences between rivers and lakes (i.e., type I errors). The purpose of this comment is to reevaluate the conclusions of Randall et al. in light of spatial autocorrelation. We accomplish this by examining their data for spatial patterns and then using two complementary analytical approaches for comparing fish attributes between rivers and lakes that are appropriate for spatially autocorrelated data. We have other concerns about their study, in particular, their sampling design is not globally well balanced; streams were clustered in northern Europe and the south Pacific, whereas no lakes were from these areas but were instead largely from Canada, Scandinavia, and Russia. There are no analytical approaches, including ones we suggest, that will remedy all biases in interpretation that may be caused by this sampling design. Further, some of the variables used by Randall et al. (e.g., production, biomass, weight, and P/B) are not independent of one another (an issue that has been discussed by Jackson et al. (1994)), yet were treated as independent variables for analytical and interpretative purposes. Moreover, field sampling in rivers and lakes requires different methods with different efficiencies that could bias their contrasts. However, we will not deal here with the sampling biases and independence issues listed above. In our comment, we focus exclusively on the effects of spatial autocorrelation because there still seems to be a general misunderstanding or ignorance among aquatic ecologists about statistical approaches that are appropriate for broad geographic scale survey data. This is despite the publication over the past 10 years of papers that detail the potential pitfalls of spatial autocorrelation in ecological studies (e.g., Legendre and Troussellier 1988; Legendre and Fortin 1989; Legendre 1993; Hinch et al. 1994; Dunham and Rieman 1999). Spatial autocorrelation may be the result of important underlying environmental gradients and it is thus critical to consider such patterns explicitly. However, frequently, one is unaware of such gradients or the scale at which they influence patterns in nature, and autocorrelation analysis can at least expose the presence of such patterns. We are concerned here primarily with the biases that autocorrelation may impose on statistical inferences and not whether important gradients may be ignored as a consequence.