Data integration for large scale ecological models

点击翻译按钮获取中文摘要

Ecological models are becoming larger, more complicated, and being used for an increasingly wide range of applications, from describing trends and mapping distributions to understanding mechanistic relationships and predicting the impact of future scenarios. In response, there has been a huge growth in statistical methods for large-scale ecological models. However, most such methods do not account for the fact that ecological data is inherently heterogeneous, and large datasets typically contain many forms of bias.Recently, a set of hierarchical Bayesian models (HBMs) have emerged as promising ways for dealing with biased data, particularly for occurrence records and other unstructured data. Many millions of unstructured occurrence records exist, so the potential of these new methods is enormous. Not all data contain biases, though. A minority of biodiversity data is highly structured in terms of the sample locations, fixed protocols and regular sampling. Ideally, we'd like to retain the information about this in our models, but combine it with the much larger sample sizes of unstructured datasets.Integrated models provide a way to do this. They are a subclass of HBM in which data heterogeneity is modelled explicitly, by treating datasets with different observation processes as independent realisations of the same underlying state. For example, causal observations on GBIF and the Breeding Bird Survey both contain information about whether the population of a particular species was extant at a particular point in space and time.At present, these integrated models are the preserve of highly competent statisticians. They are hard to specify and difficult to fit and diagnose. One goal of this partnership is to build an extensible framework for fitting integrated models that will make them accessible to a broad community of ecological modellers. This framework, in the form of open source tools, will make it easier for ecologists to handle biased data when addressing large-scale questions about biodiversity.Although attractive from a conceptual standpoint, it is unclear whether the sophistication of integrated models deliver real benefits over simple ones. In particular there is an urgent need for some general principles about how to proceed when both structured and unstructured data sources are available. Critical questions include:Q1. When and how should we combine datasets with different properties?Q2. Under what circumstances is simple aggregation (i.e. ignoring the different observation processes) better than integration? Q3. If we suspect the data contain biases, can we detect them and handle them adequately?Q4. What are the most appropriate metrics for information content and model fit?These general questions lie at the intersection of the research interests of PI Isaac, Co-I Henrys and Project Partner O'Hara. Each has made some progress towards addressing specific aspects of these questions. Working in partnership would add significant value to each, by taking existing research beyond the specific context and toward general answers to these big questions. It would permit a co-ordinated effort and build a work program of international significance. This pump-priming award would provide a platform for this partnership. The overall aim is to build a framework for inference in large-scale models of species' distribution, and to test it using computer simulations.