Talk:Maximum spacing estimation/GA1

GA Review

Article (edit | visual edit | history) · Article talk (edit | history) · Watch
Preliminary review: I think this is within reach of being a "good article" but could be improved in several ways.

• It would be good to start by explaining the concept in words before giving a formal definition, i.e., saying that the method fits parameters by choosing values that cause the values of the cum. dist. func. at the observed data points to be as evenly spaced as possible.
  • Valid point, I'll try to get to it over the next few days. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    •  Done -- Avi (talk) 02:43, 14 January 2009 (UTC)
      • I added a bit more along these lines to the lead -- feel free to edit, or course. Looie496 (talk) 18:00, 14 January 2009 (UTC)
• The formal definition is too terse: the wikilinking makes it hard to understand. For example, "distibution" hides "cumulative distribution function", and I first read "random ordered" as "randomly ordered", which caused me great confusion. I suggest converting the definition to 2-3 understandable sentences instead of one sentence that has 7 wikilinked terms.
  • Valid point, I'll try to get to it over the next few days. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    •  Done, but the formal nature of the definition should be maintained. I think this is a decent compromise. Thoughts? -- Avi (talk) 02:54, 14 January 2009 (UTC)
      • Looks good; I made one more slight tweak. Looie496 (talk) 18:00, 14 January 2009 (UTC)
• Ought to say something about robustness if possible: do bad things happen if you try to fit one type of distribution to data that actually obey a different type?
  • Your predictions may be off, but that is a matter for general statistical analysis, I believe, not this one particular type. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    •  Not done That is referenced in the goodness of fit section where one of the benefits of the method is if the Moran's statistic is too high, that indicates that the model is not fit properly (and it would thus be rejected at higher levels of significance) so I am not certain what else should be added here -- Avi (talk) 03:03, 14 January 2009 (UTC)
      • Could clarify how a consumer would make use of this, but I consider this response satisfactory. Looie496 (talk) 18:00, 14 January 2009 (UTC)
• Ought to say something about practical usage. Has anybody ever actually used this method in practice, if so why, and how well did it work out?
  • In practice? I'm not certain we are privy to work papers, but I did come across some practitioners' papers so I think I can work something in. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    •  Done Added one-sentence third paragraph to lead with sources. -- Avi (talk) 17:15, 21 January 2009 (UTC)
• Note 2 is "original research", and it would be better if you could handle this differently, but I don't see this issue as a killer.
  • I understand. It really is there so other people, if they follow the papers, can be made aware. I've actually sent an e-mail to Dr. Cheng, about this, but that is neither here nor there. It can be removed if necessary. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    • The mentioned issues in note 2 were confirmed by Dr. Cheng to be typos, for what it is worth. -- Avi (talk) 17:16, 21 January 2009 (UTC)
• The handling of ties feels a bit weird to me. The only way you can have "true" ties is to have a discontinuity in the CDF. Does this mean that the model must support discontinuities?
  • I believe the C&A paper was talking in theory, as the possibility of having two exactly same values on a continuous distribution almost never happens. C&S discuss a method of handling limits of precision in measurement. I felt both should be addressed. -- Avi (talk) 03:56, 7 January 2009 (UTC)
    • I have been in communication with Dr. Cheng, who has been kind enough to answer some of my questions. Perhaps I'll pose this one to him. -- Avi (talk) 13:50, 14 January 2009 (UTC)
• Is there anything a user can do to test whether the method is being used appropriately?
  • Checking . People can make mistakes in algebra too. Is there something more specific you had in mind? -- Avi (talk) 03:56, 7 January 2009 (UTC)
    • I see that the Goodness of fit section addresses this. Looie496 (talk) 18:00, 14 January 2009 (UTC)

These are my initial impressions -- I will probably come up with more. Looie496 (talk) 19:41, 6 January 2009 (UTC)

Thank you very much, I'll get to these points over the next few days, i hope, and drop you a line when I have addressed them. -- Avi (talk) 03:56, 7 January 2009 (UTC)

As a former mathematician who now does experiments, I tend to view statistical methods more as a consumer than a producer. From a consumer's point of view, the most critical question is whether a method is "safe", that is, whether it is possible to be assured that it won't give completely wrong results. With real data, one generally doesn't know for certain what family the "true" distribution belongs to, so the assumptions this method is based on usually aren't met. It seems likely that the method will still give an "optimal approximation" in most cases -- perhaps the approximation that minimizes the Kolmogorov-Smirnov difference? The other aspect of "safety" that has practical importance to a consumer is to know that other people have used the method and had their results accepted -- that's why I asked for examples of application if you know of any. Neither of these things will keep me from passing the article, but I think the article would be more useful if they could be covered. Looie496 (talk) 17:45, 7 January 2009 (UTC)

I'll see what I can find. A bit crazy at work now, so it may take me some time to implement your suggestions. Thank you. -- Avi (talk) 14:37, 11 January 2009 (UTC)