Tue, Aug 03, 2021:On Demand
Background/Question/Methods
Where and when species persist depends on a multitude of interacting abiotic conditions and biotic relationships. Because of the profusion of parameters needed to characterize species interactions, models of coexistence are often constrained to two-species pairs or broad species groups. We introduce a third option: adapting Bayesian sparse modeling approaches of parameter shrinkage, often applied to linear models of population genetics, to non-linear models of species interactions in diverse communities. We used a version of the Beverton-Holt model of species interactions to predict community dynamics across a gradient of environmental conditions, allowing intrinsic growth rates and species-pair interactions to vary with environment. We implemented Bayesian parameter shrinkage to determine which species-pair-specific interactions should remain in the final model and which can be sufficiently represented by a generic interaction term. This allowed us to explore the effects of environmental variation on competition among species pairs while maintaining a manageable set of parameters. We applied our model to simulated data of 15-species communities perturbed from equilibrium, and to field data from Jasper Ridge grassland communities. We tested our model's predictive accuracy with different sample sizes and on communities of varying complexity.
Results/Conclusions With as few as ten input data points, our model was highly successful at predicting out-of-sample population dynamics, with a correlation ρ of 0.86 between the predicted and observed growth rates, and average 95% credible interval width around the predicted growth rate of 0.19. Increasing the sample size to 50 improved the model accuracy to ρ of 0.94 and precision to CI width of 0.13; additional data points up to a sample size of 200 increased precision but not accuracy, likely due to underlying stochasticity in our simulation model. Increasing the sample size to 50 also improved model fit of the underlying parameters, including selecting more species as non-generic competitors, and improving the precision of the estimation of growth rate and competition parameters. In our empirical analyses, we found environmentally-dependent and species-specific growth rate and interaction coefficients. Our results demonstrate that sparse modeling approaches have tremendous potential value in analyzing interacting biotic and abiotic factors in diverse ecological communities.
Results/Conclusions With as few as ten input data points, our model was highly successful at predicting out-of-sample population dynamics, with a correlation ρ of 0.86 between the predicted and observed growth rates, and average 95% credible interval width around the predicted growth rate of 0.19. Increasing the sample size to 50 improved the model accuracy to ρ of 0.94 and precision to CI width of 0.13; additional data points up to a sample size of 200 increased precision but not accuracy, likely due to underlying stochasticity in our simulation model. Increasing the sample size to 50 also improved model fit of the underlying parameters, including selecting more species as non-generic competitors, and improving the precision of the estimation of growth rate and competition parameters. In our empirical analyses, we found environmentally-dependent and species-specific growth rate and interaction coefficients. Our results demonstrate that sparse modeling approaches have tremendous potential value in analyzing interacting biotic and abiotic factors in diverse ecological communities.