User talk:SolidPhase

Disambiguation link notification for February 27
Hi. Thank you for your recent edits. Wikipedia appreciates your help. We noticed though that when you edited Generalized linear mixed model, you added a link pointing to the disambiguation page Springer. Such links are almost always unintended, since a disambiguation page is merely a list of "Did you mean..." article titles. Read the FAQ* Join us at the DPL WikiProject.

It's OK to remove this message. Also, to stop receiving these messages, follow these opt-out instructions. Thanks, DPL bot (talk) 09:18, 27 February 2015 (UTC)


 * Fixed. Thank you, bot.   SolidPhase (talk) 09:50, 3 March 2015 (UTC)

Re Grey box model
Thanks for the work done on the page “Grey box model” and also your kind comment on the potential usefulness of the page. An not sure how the reference to “Model theory” crept in, but the title for that page might be better as “Model theory (logic)’. I will consider extending the content now there is a more general title.

As you noted the work is not well known and I think it would be helpful if there were additional references to the page, which had sat as an orphan for over a year. I have answered your query regarding conflict of interest on my talk page. I note also that software is available in the public domain. Last time I checked there had been no other objections to the “See also” references, and there was one thank you.

So if you are agreeable I would like to restore the “See also” references so that the page is more readily found.

The three references you removed actually contain a lot information on grey box modelling techniques that is too detailed for inclusion on the page. The two thesis references are publicly available. I was originally asked to ensure the article was extensibly supplied with references. SoI would like to reinstate those references.

Kind regards, BillWhiten (talk) 10:15, 5 April 2015 (UTC)


 * I have restored the three "See also" references.
 * Are the two thesis references readily available to readers in the U.S. and U.K.? Are there web links? If not, would it be feasible to put the thesis references online somewhere, and then link to them? Is a 1971 reference really still useful?
 * Something that I am not clear about is whether a grey box model must always also be a statistical model. The first paragraph from Statistical model is relevant here, and is copied below.
 * "A statistical model is a special type of mathematical model. What distinguishes a statistical model from other mathematical models is that a statistical model is non-deterministic. Thus, in a statistical model specified via mathematical equations, some of the variables do not have specific values, but instead have probability distributions; i.e. some of the variables are stochastic."
 * In the equations of a grey box model, is it required that there be stochastic variables? I think that the article would benefit from making that clear.
 * SolidPhase (talk) 10:59, 5 April 2015 (UTC)

Thesis are available through library exchanges or purchase. Not sure about on line. I do not have copyright to put on line. Grey box models can be stochastic or deterministic depending on how user sets up and uses model, have changed text to indicate models can be mathematical ,statistical, or computational. BillWhiten (talk) 10:19, 9 April 2015 (UTC)


 * Your citing of your own work is a violation of WP:CoI. Additionally, I a skeptical of citing a work from 1971: surely there must be superseding references available. (Ideally, those would be references that are easily obtained.) Hence, I ask you to remove the 1971 reference. SolidPhase (talk) 14:10, 16 April 2015 (UTC)

Your cleanup statistics, Akaike information criterion
Hi SolidPhase. I wouldn't want to doubt your good intentions, but are you absolutely certain that simple erasure of information (esp. in former paragraph AIC "Limitations" - I think I heard they actually do exist) is really beneficial? As for the lead section of the statistics article, keeping the crucial keywords like "confidence interval" along with a short sentence and a link to the relevant paragraph below could have been an alternative. Purging the mention of "big data" is no big saver and deprives the last sentence of any concreteness. -- Kku 10:29, 19 June 2015 (UTC)
 * Hi Kku. The claim made in the section added by 93.24.5.93 was misleading. The so-called limitation is that AIC does not work well when n is small: this issue is explained in detail, with references. So, yes, the change I made was beneficial.
 * About Statistics, see WP:LEAD, which was cited in my edit summary: see especially the sections entitled "Length" and "Clutter". The keywords that you mention could be confusing for readers who do not know them, and unneeded for readers who do. Regarding "Big data", I do not really agree, but okay I put it back in.
 * SolidPhase (talk) 10:44, 19 June 2015 (UTC)

ArbCom elections are now open!
Hi, You appear to be eligible to vote in the current Arbitration Committee election. The Arbitration Committee is the panel of editors responsible for conducting the Wikipedia arbitration process. It has the authority to enact binding solutions for disputes between editors, primarily related to serious behavioural issues that the community has been unable to resolve. This includes the ability to impose site bans, topic bans, editing restrictions, and other measures needed to maintain our editing environment. The arbitration policy describes the Committee's roles and responsibilities in greater detail. If you wish to participate, you are welcome to review the candidates' statements and submit your choices on the voting page. For the Election committee, MediaWiki message delivery (talk) 17:07, 24 November 2015 (UTC)

AIC problems
I note that you did not allow me to make changes or contribute to the Akaike information criterion by undoing my efforts. AIC relies on the assumption that goodness-of-fit is the only criterion for model selection. That assertion is false. Models are not only used to create biased estimators of functional amplitude at the fit points, which is your assumption. In specific, one man's information is another man's noise, and there is no information without context. In specific, by assuming a context, which in the case of AIC is goodness-of-fit, all other contexts are excluded, and those include regression to find the best functional relationship between variables as opposed to a least error estimation of amplitude, best accuracy of extrapolation, least biased estimator, and many many others. AIC is not as useful as the article portrays. It is not a fall back position for comparing models in any sort of general context. I would ask you to consider moderation on this point rather than just erasing my comments. I am willing to compromise, however, the article as it stands in not true, optimistic yes, true not. Also note, "If the assumptions we made in Section 2 are accepted, Akaike's criterion cannot be asymptotically optimal. This would contradict any proof of its optimality, but no such proof seems to have been published, and the heuristics of Akaike [1] and of Tong [4] do not seem to lead to any such proof." Since this has been cited 26907 times on Google Scholar, I take that as a sign that many authors have noted that AIC is not on firm ground.

CarlWesolowski (talk) 17:51, 17 May 2016 (UTC)
 * The paper of Schwarz [1978] introduced the Bayesian information criterion. There already is a subsection of the Wikipedia article that compares BIC and AIC: Akaike_information_criterion. That subsection cites five references, all published during this century.


 * Regarding the purposes of a model, the article Statistical model lists three. Additionally, see the article Model selection (which needs expansion).
 * SolidPhase (talk) 09:16, 18 May 2016 (UTC)

Bayesian information criterion is a systematization of preconceptions as if they were predictive. As such it is no more valid than any other confirmation bias. This establishes a tautology of thought that is not testable from within Bayesian argumentation. The argument reduces to the premise that the probability of a future event is determined by experience. I counter with Black swan theory, and note that it is bad form to assume an arbitrary argument and then prove that that is indeed what was assumed. The relationship between AIC and BIC is close enough to discount both as presumptuous. There are no confidence intervals for AIC, and it is not a standard test. AIC has no meaning for any model, and to say then that it means something when comparing models is based upon an information theory that is far too limited to be generally applicable. Granted, it may have some limited application for ordinary least squares in y, but ordinary least squares in y, is the tip of a very large iceberg filled with numerical methods, which methods should be tailored to the information sought during regression rather than just using OLS. I use OLS only about 20% of the time for my regression analyses, and never use it for primary event models because the frequency of occurrence of the presumed error structure is too infrequent for me to have ever not found a more appropriate statistic for minimization. Quantification of the information content maybe could be extended to other types of information. However, what is missing is any motive for doing so. For example, one would have a very hard time proving that AIC is in any way superior to adjusted R-squared, Chi-squared testing and the other information that you deleted. I admire your faith in AIC, but need help. How do I get the dangerous and misleading overly rosy presentation of AIC pruned so that it is not presented as if it were the "GO TO" method for everything?

I do not care that the other sections of Wikipedia you refer to have errors of omission, logic and so forth. They are perhaps not terribly useful, but, they are not dangerous in the same way that AIC is. AIC currently is presented more as an expression of belief than of science, and it is dangerous precisely because it represents an impediment to understanding, and I get reviewers who say "show me with AIC" for error propagation problems totally unrelated to goodness-of-fit. Thus, I see this problem constantly. I take it that you are not familiar with inverse problems as that information was deleted. The general method of solving anything is the inverse problem treatment, not the Bayesian inference, and it is by using inverse methods that we could discover what the Bayesian inference should have been, that is, if we care to do so, because it leads nowhere special. So, as I sit here among my black swans, I humbly plead with you to reconsider some of your deletions so that we can get the AIC article into a more useful form.


 * If you believe that the Wikipedia article has an error, then find the error and cite WP:RELIABLE sources to back up your finding. Thus far, you have only cited one reliable source: a paper from 1978, on a topic that already has a subsection devoted to it&mdash;a subsection which discusses the assumptions of the 1978 paper.
 * SolidPhase (talk) 16:35, 18 May 2016 (UTC)


 * A source that explicitly discusses the uses and limitations of the AIC may also help. Huon (talk) 19:38, 18 May 2016 (UTC)

If you want, I will write an article on the inappropriate uses of AIC, that is, if I cannot find that particular needle in a haystack. I will look for one, but, I think it may be hard to find since the users group into those who cannot be talked out of using AIC and those who cannot be talked into it. Failing that you might realize that you are not stating what the assumptions are for using AIC. OK let's start there. You are assuming OLS and "Results on the bias and inconsistency of ordinary least squares for the linear probability model" Portray this as not useful. Similarly, I found this not useful for gamma variate modeling. "Limitations of ordinary least squares fitting of gamma variate functions to plasma-clearance curves". The uses of weighting factors to get correct answers is standard and very common, but what it is, is not OLS. For example, weighting for relative error or minimizing the norm of model/y-1 is not OLS, and is more appropriate in some circumstances for example for measuring "Plasma concentration half-life". Similarly performing MLR on the logarithms of data will yield better models for proportional error structure than any direct OLS approach "An improved method for determining renal sufficiency using volume of distribution and weight from bolus Tc-99m-DTPA, two blood sample, paediatric data". Tikhonov regularization is definitely not an AIC problem type when the smoothing factor is greater than zero. An example of this that has nothing to do with goodness of fit is adaptively minimized relative plasma clearance error propagation using Tikhonov regularization "Method for evaluating renal function". The more general theory of inverse problem solutions can, when applied, give evidence of actual as opposed to assumed Bayesian prior information is found here. Does that help? I will rummage around to see if anyone cared enough about AIC usage to write something on it. I included one before, namely "Akaike's information criterion in generalized estimating equations" but it was deleted. So promise me you won't delete material evidence, again, please? OK, one more paper. Please read this one to gain a broader perspective on this issue. "Assessing Goodness of Fit: Is Parsimony Always Desirable?" . This paper compares many many tests including AIC. None of them replace common sense according to the authors. The actual problem is that common sense is to specify the desired measurement ahead of time and to choose methods that minimize errors in that context, and, parsimony in that context is not generally goodness-of-fit, sometimes perhaps, but not generally. — Preceding unsigned comment added by CarlWesolowski (talk • contribs) 20:03, 18 May 2016 (UTC)


 * I looked at your first reference, Horrace & Oaxaca [2006]. It does not even mention AIC; hence it has no relevance. You must cite reliable sources that make the points. You cannot violate WP:ORIGINAL.
 * SolidPhase (talk) 21:31, 18 May 2016 (UTC)


 * The first reference relates to heteroscedasticity of OLS residuals for any probability model of the worst kind, bias over the range. Least squares in y is biased when there is any randomness in x. AIC assumes goodness of fit in a maximum likelihood context, if the likelihood itself is a biased function, then the comparison between models is between incorrect and incorrect, at least I think so. To be honest, I do not understand the assumptions being made, nor where one can go from those assumptions, which seem too mushy to yield anything. What the references do is attempt to follow those assumptions out, and they seem to lead nowhere for most physical problems.

Here is a simple example: Suppose that I own shares in a company. Does goodness of fit or a likelihood analysis tell me what those shares will be worth tomorrow? Doubtful, isn't it? Suppose I persist and try to make a prediction, do I not have to match the derivatives of the trending data as well as the data itself? In fact, I do not then care if the model is parsimonious today, I only care about parsimony for tomorrow. Does AIC apply to selecting a model for tomorrow's share price? Would not extrapolation testing be a better approach to model selection, as in "see what works"?

I am trying to see what AIC is actually useful for, and so far is seems like it is seldom useful. I do not think Akaike information criterion is a balanced presentation, it looks more like an advertisement. In my opinion that is a disservice. For example, BIC has many advantages over AIC and only one side of this is presented. In the mean time, I will try to figure this out elsewhere, and you can follow along if you wish at. After much thought, I provided reasonable edits to your introduction. You should not present AIC as if it were the sole deciding factor for model selection. Attached is example output from the FindDistribution routine in Mathematica. What I want you to notice is that the following output of the top 5 of hundreds of distribution tried automatically including additive mixtures of independent distributions has no really good model, and no single criterion can lead one to that conclusion. If one chooses AIC as THE criterion then one would choose a distribution that is unlikely by Cramer Von Mises probability to represent anything significant. Without also having the distribution fit agree by preferably both Chi-Square and Cramer Von Mises probabilities with the data, then one is promoting an unlikely model indeed. I would suggest you read up on model selection before you make blanket statments about how wonderful AIC is. Yes, AIC is one tool in the box. No, it is by no means a stand-alone model selection criterion.

CarlWesolowski (talk) 04:21, 22 May 2016 (UTC)

Disambiguation link notification for May 20
Hi. Thank you for your recent edits. Wikipedia appreciates your help. We noticed though that when you edited Roger L. Simon, you added a link pointing to the disambiguation page City Journal. Such links are almost always unintended, since a disambiguation page is merely a list of "Did you mean..." article titles. Read the FAQ* Join us at the DPL WikiProject.

It's OK to remove this message. Also, to stop receiving these messages, follow these opt-out instructions. Thanks, DPL bot (talk) 09:54, 20 May 2016 (UTC)
 * Thank you&mdash;I have revised the link. SolidPhase (talk) 10:31, 20 May 2016 (UTC)

Discretionary sanctions alert, please read
Doug Weller talk 14:31, 27 April 2019 (UTC)

Removal of reference from ordinary least squares
Your removal of Davidson/MacKinnon created an incomplete reference, since three footnotes (#28 to #30) referred to the book in Harvard citation. Please be more careful when your remove references. --bender235 (talk) 23:28, 9 June 2019 (UTC)