Challenging and diagnosing structured population models by testing predictions from stochastic demography

1. Structured population models are parameterized to accurately project expected population sizes, stage/state distributions and population growth rates, but they also predict the variation in outcomes among individuals, such as the variance and skewness of lifetime reproductive output (LRO) and lifespan, the probability of never reproducing, and many other life-history metrics.

2. Testing such predictions about individual outcomes can be very useful model ‘stress tests’, because they depend on how components of the model (e.g. submodels for survival and fecundity) interact over multiple time steps, not just on the accuracy of the submodels. Because data on among-individual variation is rarely used to parameterize the models, models will not automatically pass the tests.

3. We present case studies (including zooplankton, plants and mammals) to demonstrate how structured population models can be tested by comparing individual-level predictions from existing models against individual-level data.

4. Some general themes emerge: (i) We often detect unmodelled individual heterogeneity, (ii) Unmodelled senescence can affect higher moments of lifespan even when lower moments and LRO are predicted well. (iii) Fitting one parametric model to multiple clones, species, locations, etc. can lead to poor predictions about groups for which the model is insufficiently flexible.

5. The ways in which model predictions fail can help to identify what the problems are, help us decide whether the problems are important for the model's intended purpose, and guide efforts to fix them. Structured population models are ‘workhorses’ for ecology: tests based on predictions from stochastic demography can help ensure their reliability.