Another potentially problematic issue with model averaging is the selection of candidate probability models. As noted by Burnham and Anderson (2002), the construction of the candidate model set involves an element of subjectivity and that “one must recognize a certain … Read More
It has been suggested that model averaging is only likely to be useful when (a) the uncertainty in contributing model estimates is dominated by the conditional uncertainty of each model and not among-model uncertainty; and (b) the covariance between models … Read More
As noted by Fox (2016), small samples result in low-powered goodness-of-fit tests meaning any plausible candidate model is unlikely to be rejected by these procedures. It is somewhat puzzling that the cite Chapman et al (2007) but do not note … Read More
It is not clear how the ability to edit open source code enhances harmonization of approaches. It seems reasonable to think new approaches or modifications could ultimately improve the science by demonstrating improvements, but it also seems clear that, at … Read More
The authors express strong opposition to the use of compiled code, such as in the recent SSD Toolbox released by USEPA. So long as the methodology and algorithms are clearly documented, their opposition to compiled code seems misguided. The use … Read More
The paper also considers future directions both with and beyond the use of SSDs, with the ultimate aim of helping to facilitate greater international collaboration and, potentially, greater harmonisation of SSD methods. Given that aim, it is interesting that in … Read More
Mixture data is illustrated as a general idea, but poorly illustrated in the context of SSDs and not well motivated as to relevance.
The novel sections are the advocacy of model averaging and mixture models. The model averaging section should be expanded to include some noted cautions (discussed below) and more details on how the two highlighted software packages implement this. At least … Read More
The list of software is interesting, but they do almost nothing with it. There is no useful discussion of what these packages do, how to use them, their effectiveness or limitations. Either develop this, put it in supplementary material, or … Read More
The authors should decide what paper they want to write. The initial history of SSDs with references is good, but much too brief for a review paper and omits one important topic (censored data) and barely mentions another (Bayesian methods).