On the setting of environmental noise and the performance of population dynamical models

Abstract

BackgroundEnvironmental noise is ubiquitous in population growth processes, with a well acknowledged potential to affect populations regardless of their sizes. It therefore deserves consideration in population dynamics modelling. The usual approach to incorporating noise into population dynamical models is to make some model parameter(s) (typically the growth rate, the carrying capacity, or both) stochastic and responsive to environment fluctuations. It is however still unclear whether including noise in one or/and another parameter makes a difference to the model performance. Here we investigated this issue with a focus on model fit and predictive accuracy. To do this, we developed three population dynamical models of the Ricker type with the noise included in the growth rate (Model 1), in the carrying capacity (Model 2), and in both (Model 3). We generated several population time series under each model, and used a Bayesian approach to fit the three models to the simulated data. We then compared the model performances in fitting to the data and in forecasting future observations.ResultsWhen the mean intrinsic growth rate, r, in the data was low, the three models had roughly comparable performances, irrespective of the true model and the level of noise. As r increased, Models 1 performed best on data generated from it, and Model 3 tended to perform best on data generated from either Models 2 or Model 3. Model 2 was uniformly outcompeted by the other two models, regardless of the true model and the level of noise. The correlation between the deviance information criterion (DIC) and the mean square error (MSE) used respectively as measure of fit and predictive accuracy was broadly positive.ConclusionOur results suggested that the way environmental noise is incorporated into a population dynamical model may profoundly affect its performance. Overall, we found that including noise in one or/and another parameter does not matter as long as the mean intrinsic growth rate, r, is low. As r increased, however, the three models performed differently. Models 1 and 3 broadly outperformed Model 2, the first having the advantage of being simple and more computationally tractable. A comforting result emerging from our analysis is the broad positive correlation between MSEs and DICs, suggesting that the latter may also be informative about the predictive performance of a model.