Extended Data Fig. 6: A total of 64 different plausible network models were assessed using expected log pointwise predictive density (ELPD) and leave-one-out (LOO) cross-validation for predictive validation.

(a) The figure contains the best network, Model 36 (top), plus all of the other networks within 10 leave-one-out information criteria (LOOIC) units of Model 36. The node labels are abbreviated; spill is spillover, roost is roost fissioning, land use is the land use type at the roost level, short is a food shortage in the previous year, pulse is a winter flowering pulse, and ONI is elevated ONI over 0.8 two years prior. (b) Distribution of LOOIC values relative to Model 36. For each model, the boxplots (n = 1000 simulations) show the distribution of the difference in LOOIC. M36 is the favored model as the median, and at least the first quartile, are positive for all other models. The bounds of the box correspond to the 25^th and 75^th percentiles, the middle dash is the median, and the whiskers extend to the largest value no further than 1.5 times the interquartile range from the 25^th or 75^th percentile.