Sub-optimal study design has major impacts on landscape-scale inference

Landscape-scale processes (e.g. habitat loss) are major drivers of the global biodiversity crisis, but the complexity and size of landscapes makes study design at this scale difficult. However, the impact of statistical problems associated with sub-optimal study design on inferences drawn from landscape-scale studies is poorly understood. Here, we examine how three common statistical 'pitfalls' associated with sub-optimal study design - (1) using landscapes that overlap in space; (2) using only a portion of the potential range of the landscape predictor variable(s) of interest; (3) failing to account for correlations among landscape predictor variables - affect the inferred relationships between the abundances of six species of anurans and the amount of forest in the landscape using a large (n = 1141) empirical dataset from Wisconsin and Michigan, USA. We show that sub-optimal study design alone can be sufficient to cause a switch in the sign of the inferred relationship between a species response and landscape structure, and that using only a portion of the potential range of a predictor variable, and correlations between predictor variables, are particularly likely to affect inferences. Our results also provide the first evidence of a non-monotonic relationship between forest amount and gray treefrog abundance, and suggest that inconsistencies in the literature about the inferred relationships between anuran presence/abundance and forest amount in the Great Lakes basin are likely largely due to sampling design issues. Increased attention to study design is therefore necessary for the development of robust generalizations in landscape ecology.