Measuring metrics: what diversity indicators are most appropriate for different forms of data bias?

Biodiversity metrics have become a ubiquitous component of conservation assessments across scales. However, whilst indices have become increasingly widely used, their ability to perform in the face of different biases has remained largely untested under realistic conditions. Citizen science data are increasingly available, but present new challenges and biases, thus understanding how to use them effectively is essential. Here, we built a virtual world incorporating BirdLife data and accounting for their biases, then explored how well commonly‐used diversity metrics could estimate known values across a suite of representative scenarios. We used predictive modelling to model bird diversity globally and overcome biases using the approaches found most accurate in prior assessments. Performance was highly variable across the different types of biases, but in many instances Simpson's index performed best, followed by Hill numbers, whereas Pielou's index was almost universally worst. From standardised tests, we then applied these metrics to eBird data using 611 520 112 samples of 10 359 species of bird (around 88% of known species), to reconstruct global diversity patterns at five and ten km resolutions. However, when we mapped out diversity using Maxent based on these indices, Simpson's index generally over‐predicted diversity, whereas Hill numbers were more conservative. Based on an average of the better projected indices, one can map out diversity across resolutions and overcome biases accurately predicting diversity patterns even for data‐poor areas, but if a single metric is used, Hill numbers are most robust to bias. Going forward, this workflow will enable standardized best practices for diversity mapping based on a clear understanding of the performance of different metrics.