Put comments here for "Ecosystem analysis using probabilistic relational modeling" by Bruce D'Ambrosio, Eric Altendorf, and Jane Jorgensen
From James Cussens: I'm surprised that data analysis methods are not well developed for ecosystems, since the study of ecosystems is not new. Is there a reason for this (possibly something sociological?)
Am I right in thinking that the creation of synthetic variables is a form of propositionalisation?
From Bruce D'Ambrosio: There are two largely disconnected families of modeling efforts in ecosystems. Deterministic first-principles models often are cross-disciplinary (although rarely cross-scale), for example global climate models. However, these only loosely touch data for model verification purposes. Mostly verification is intra-disciplinary (ie, piecewise). Modeling efforts that start from data tend to focus on single disciplines. For example, a previous several hundred page report on the Crater Lake data used in our report restricted itself to single variable and two-way, intra-disciplinary, analyses. Prior to relational methods, there were no methods available for cross-disciplinary investigation.
synthetic variables = propositionalization? Yes and No. In the fully observed case, yes, it is equivalent. However, when there is missing data or hidden variables, then the scope (table) in which a variable exists is crucial in both structure discovery and parameter estimation (e.g., (I'm playing rather loose with d-separation here to be brief) if two people share a parent, then if the parent's age is known, the sharing is irrelevant, but if the parent's age is unknown, there is only one instance, and it ties the two children together.)