Notice: Undefined index: img in /home/environ5/public_html/wp-content/plugins/wp-members/includes/class-wp-members-forms.php on line 1144

Notice: Undefined index: hidden in /home/environ5/public_html/wp-content/plugins/wp-members/includes/class-wp-members-forms.php on line 1144

This content is restricted to site members. If you are an existing user, please log in. New users may register below.

Existing Users Log In
   
New User Registration
Really Simple CAPTCHA is not enabled
*Required field
3 Comments
Newest
Oldest
Inline Feedbacks
View all comments
David Fox

RESOLUTION

No change. Make following points in response:

1.      this (Numerical stability) and other issues listed are in the Future Directions section and this was more a note to self in which we identify what work needs to be done. As we say in the MS “In terms of our own R&D efforts, we have identified the following priority issues that will form the basis of further collaboration between Australian and Canadian jurisdictions”. So it was never intended to provide full treatment of topic and identify solutions – that’s what we’re saying needs to be done!

2.      Given that a number of us have been instrumental in the development and integration of Burr/Burrlioz into ecotox over ~25 years, I think we can rightly criticise on the basis of this long experience (albeit a “brief criticism”).

Carl Schwarz

I think the major issue for numerical stability is that you want your software to be robust. Most users of software are not statisticians and so don’t know how to evaluate a fit (e.g. non-sensical standard errors, non-convergence) and assume that if the software runs, the output must be correct. So good algorithms are needed to ensure that the software converges even under extreme cases (e.g. sample size of 4!) that we cannot predict in advance. The software should rely on the user checking the reported vcv matrix and or the curvature of the likelihood surface at the final convergence point (e.g. after 100 iterations when it stops due to a very flat likelihood).

David Fox

The reviewer’s statement “the brief criticism given is not helpful” is bewildering. It’s a statement of fact based on observation. The full sentence in our MS reads:

“While offering a high degree of flexibility, experience with these distributions during that time has repeatedly highlighted numerical stability and convergence issues when estimating parameters using maximum likelihood”.

As the person responsible for introducing the Burr family into the SSD lexicon together with colleagues GB and RvD, we have more than 60 years of accumulated experience with this distribution and the Burrlioz software. On this basis, I think we are well-placed to make such commentary (criticisms?)!!

Further, and to the issue at hand: I (we?) agree with the statement “Every distribution used for SSDs has limitations”. That is completely understood. But for me these limitations are of varying complexity and consequence. A disrtibution like the log-logistic is, for the most part, well-behaved. Any limitations with this distribution in the context of SSDs is more to do with the fit provided and not the estimation of parameters. I regard the former as a second-order issue and the latter as a first-order issue. So, failure to converge or multiple optima for me are more serious (in terms of consequences) than a failure to fit.