The authors attempt to remove unintentional over-weighting of similar model by restricting candidate models to one for each “shape.” It is not clear that this is desirable or what “shape” means here. If models that have the same “shape”, whatever that means, fit the data best by some agreed-on criteria, is it wise to eliminate all but one of them from the averaging process? Does this not tend to introduce artificially inflated uncertainty in the process?

3 Responses

  • We can clarify in the revision but my response is:

    First “shape’ has both a strict and colloquial interpretation. Many of the distributions used for SSDs have a shape parameter (lognormal, loglogistic, Burr etc.). So in that sense we don’t have to provide a definition for accepted statistical terminology. In the more general context, I would have thought “shape” was a term that was well understood. In SSD modelling I would say we’re talking about: symmetry (or lack of) = skewness; peakedness = kurtosis; and how ‘heavy’ the tails are. Distributions that basically look the same (eg a normal and a logistic) will, by anyone’s assessment, have similar shapes.

Comments are closed.