Methods in Ecology and Evolution just posted a paper I wrote with my colleague Joe Clark in which we set out to address the question: How do you draw inferences about the mechanisms governing population dynamics at large spatial and temporal scales?
This is a tough problem with the crux of the issue being that a mechanistic understanding of population dynamics requires demographic data, which typically can only be collected using expensive capture-recapture methods. So how do you scale up inferences when you can’t afford to implement such costly studies over large areas or long time periods?
The solution we propose is to combine capture-recapture data with “cheaper” data, such as count data or binary detection/non-detection data, which can be collected at broad scales, and conduct a joint analysis in which both data sets are conditioned on a spatially explict model of population dynamics. The model for the population dynamics is a point process model that describes spatial and temporal variation in population density as a function of (apparent) survival, recruitment, and possibly dispersal processes. These ecological processes can be modeled as functions of individual characteristics (e.g., sex or age) as well as spatial covariates such as landscape features, so it’s possible to assess a wide range of ecological hypotheses.
The idea of combining capture-recapture data with count data has been around for a long time, and a rigrous statistical approach was first proposed by Besbeas et al. 2002. The class of models they developed has come to be known as “integrated population models”, and a good review of the approach is given by Schaub and Abadi (2011) and in chapter 11 of Kéry and Schaub (2012). Our work represents a spatially explicit extension that resolves some of the shortcomings of non-spatial models. For instance, our model can be used to predict abundance in unsampled areas, at any spatial scale, and it does not require that the two data sets be independent.
While I think this approach offers tremendous promise for scaling up inferences about population dynamics, it is computationally demanding, and so far, it has only been applied to data from a single study. I’m eager to hear how well it works in other contexts, so if you give it a try, please let me know.