COS 313-1 - Can we identify key species interactions from observational time series?

Abstract: Parameterizing models of ecological community dynamics can help inform management decisions by providing a means of forecasting ecological states. However, most models of community dynamics are parameter-rich and include matrices of pairwise interaction coefficients to model the effects of biotic interactions. Thus, fitting these models requires quantities of data that may not always be practical. Alternatively, there is mounting evidence to support that many ecological communities may be effectively described by "reduced" models with interaction matrices that lie somewhere between a fully parameterized matrix that includes unique interaction coefficients for each species pair and one with a single common coefficient for all species pairs. But, which interactions should be collapsed to an average value and which should be estimated separately? Bayesian learning techniques have shown promise for unsupervised model selection in this context, but so far only using data collected from intensive experiments in which membership and abundance in replicate communities is manipulated. We asked whether Bayesian learning can be used to fit reduced models of community dynamics using observations of population growth through time.

To understand the feasibility and limitations of this approach, we simulated competitive community dynamics using a Ricker competition model with log-Normally distributed errors in which species had strong competitors that made up a subset of the full community while the competitive effects of others were negligible. We then asked whether we could use Bayesian sparse learning methods to identify the strong competitors of an arbitrarily selected focal species assuming we have observed abundances of each species in the community over a 50-year time series. In general, we found that there was little information on competition in our simulated observational time series, particularly when sources of error, such as demographic stochasticity, had large variance. Our baseline true positive rate (TPR), or the ability to detect strong competitors out of the background of weakly competing species, was 0.002 (95% confidence interval [5.07×10^-5, 0.01]) over 200 simulated communities. However, we found we could increase the TPR to 0.63 [0.58, 0.67] without a sacrifice in accuracy (0.97 [0.96, 0.98]) by increasing the mean temporal variance of heterospecific abundances by a factor of 10. These results suggest that simple experiments to amplify the variance of heterospecific abundances through time, such as thinning experiments, may allow researchers to parameterize reduced community dynamics models. Reduced models may be useful for the sake of forecasting or for prioritizing resources in future detailed experiments.