Effects of population-level collection biases on model accuracy
Phenoclimate models derived from sample data that exhibited population-level biases in collection timing (i.e., biases towards collection of early or late individuals from within each population) exhibited substantially higher MAE than models produced using unbiased collections (Fig. 4). The greatest increases in MAE were observed among predictions of flowering termination derived from collections biased towards early-flowering individuals, and among predictions of flowering onset derived from collections biased towards late-flowering individuals. However, predictions of median (peak) flowering DOY derived from early- or late-biased collections also exhibited significant reductions in accuracy, with mean MAE among models derived from biased collections sometimes exceeding two weeks (16.4 Days, Fig. 4). Moreover, phenoclimate models appear to be highly sensitive to < 2 days of population-level temporal biases in collections, with MAE of all predictions exceeding 5 days even when skew was low (α = ± 0.25, Table S3, Figs. S1b, S3), more than doubling the observed MAE of phenoclimate models developed from unbiased collections.