Optimizing sampling approaches along ecological gradients

Summary Natural scientists and especially ecologists use manipulative experiments or field observations along gradients to differentiate patterns driven by processes from those caused by random noise. A well-conceived sampling design is essential for identifying, analysing and reporting underlying patterns in a statistically solid and reproducible manner, given the normal restrictions in labour, time and money. However, a technical guideline about an adequate sampling design to maximize prediction success under restricted resources is lacking. This study aims at developing such a solid and reproducible guideline for sampling along gradients in all fields of ecology and science in general. We conducted simulations with artificial data for five common response types known in ecology, each represented by a simple function (no response, linear, exponential, symmetric unimodal and asymmetric unimodal). In the simulations, we accounted for different levels of random and systematic error, the two sources of noise in ecological data. We quantified prediction success for varying total sample size, number of locations sampled along a spatial/temporal gradient and number of replicates per sampled location. The number of replicates becomes more important with increasing random error, whereas replicates become less relevant for a systematic error bigger than 20% of total variation. Thus, if high levels of systematic error are indicated or expected (e.g. in field studies with spatial autocorrelation, unaccountable additional environmental drivers or population clustering), continuous sampling with little to no replication is recommended. In contrast, sampling designs with replications are recommended in studies that can control for systematic errors. In a setting that is characteristic for ecological experiments and field studies strictly controlling for undeterminable systematic error (random error ≥10% and systematic error ≤10% of total variation), prediction success was best for an intermediate number of sampled locations along the gradient (10–15) and a low number of replicates per location (3). Our findings from reproducible, statistical simulations will help design appropriate and efficient sampling approaches and avoid erroneous conclusions based on studies with flawed sampling design, which is currently one of the main targets of public criticism against science.