Effects of sample size on model accuracy
Although sample size exhibited significant effects on model accuracy,
the magnitude of the changes in MAE was < 2days , and remained
consistent across predictions of flowering onset, median, and
termination DOYs (Fig. 3). In all cases, MAE declined with larger sample
sizes, but MAE exhibited < 2 days improvements as sample size
increased from 100 to 1000 specimens in all cases (Fig. 3, Table S2).
Increases in model performance as sample size increased above 300 were
minimal, and never exceeded a one-day reduction in MAE (Table S2).
Increased magnitudes of unexplained (i.e., stochastic) variation in
phenological timing among populations within a species also were
associated with increased MAE as expected (Fig. 3), but exhibited
similar relationships to sample size as noiseless models. This implies
that unexplained phenological variation inherently degrades the accuracy
of phenoclimate models, but the effects of unexplained variation in
phenological timing cannot be remedied by greater quantities of sample
data.