Getting the errors right: The importance of partitioning sources of uncertainty for ecological indicators

•Ecological indicators should be characterized by their probability distribution.•Indicator confidence is overestimated with inappropriate uncertainty partitioning.•Variance parameters can be estimated from large monitoring data sets.•Uncertainty associated with sampling and analysis procedures are considerable.•Sampling strategies should be informed by uncertainty estimates of error sources. Environmental policies such as the Water Framework Directive (WFD) requires that the confidence of ecological status assessments should be reported. Such assessments are typically based on aggregating several ecological indicators, but the uncertainty of these is rarely quantified and when exceptionally it is estimated, the indicator variance can be grossly underestimated, resulting in overconfident assessment. We demonstrate with a simple example that incorrect partitioning of different sources of variation, typically characterizing monitoring data, can underestimate the standard error of an indicator by 33 %. This is due to sampling constraints in monitoring programs, implying that observations are not independent replicates across all levels of sampling. We also carried out a comprehensive analysis, quantifying the magnitude of different sources of variation for monitoring variables used to calculate ecological indicators for WFD status classification in Sweden. We demonstrate that these variances can be estimated from regular monitoring data, although it was not possible to estimate all relevant sources of variation. We propose to occasionally include spatial and temporal replicate samples in the existing monitoring programs such that relevant sources of uncertainty can be quantified with sufficient precision. This library of variance parameter estimates allows for calculating the uncertainty of ecological indicators more correctly for any combination of sampling in time and space by different institutions/taxonomists. It also identifies the dominant sources of random variation affecting the indicator uncertainty, providing a basis for optimal sampling design as well as potential improvement of current sampling and analysis procedures, both aiming at reducing uncertainty. Whilst it is commonly understood that sampling occasions should be spread appropriately across time and space, the relatively large variability among institutes/taxonomists implies that spreading samples among these also helps reducing indicator uncertainty, particularly for biological indices. I