Template talk:Experimental design
|WikiProject Statistics||(Rated Template-class)|
|WikiProject Mathematics||(Rated Template-class)|
- 1 Preliminary discussion of prototype
- 2 The previous discussion was about the prototype. This section begins the proper discussion of this template
- 3 Block design (Algebraic combinatorics) doesn't link to algebraic combinatorics
- 4 Comments
- 4.1 null & alternative hypothesis · type I & II error · significance · p-value
- 4.2 deduction, induction & abduction and falsifiability
- 4.3 Statistical criteria & methods: minimum-variance unbiased estimator · Gauss-Markov theorem · least squares · maximum likelihood estimation
- 4.4 ANOVA "F-distribution & test · chi-square distribution · Cochran's theorem"
- 5 Create separate Template for Linear Models and Statistical Inference?
- 6 Example of a template with fewer entries
- 7 Capitalization
- 8 Redirect of "experimental design" to this template
- 9 Edits by Melcombe
Preliminary discussion of prototype
Thanks for creation
Creating this template was a good initiative. Kudos!
Designs by factor-type (nominal versus real-number, and mixed) first, then specific designs
I would suggest having first designs by type of factor:
- nominal (e.g. categorical blocks and treatments)
- quantitative (unconstrained [usual RSM] then constrained [e.g. mixtures])
- mixed (such designs are often custom-made, of course, so there may be very few Wikipedia entries).Kiefer.Wolfowitz (talk) 12:23, 20 June 2009 (UTC)
I tried to create such an organization. For simplicity, I used "visual formatting" with line-returns, rather than creating subcategories. (I wish that this "visual formatting" would facilitate others's corrections and improvements.) Kiefer.Wolfowitz (talk) 13:47, 20 June 2009 (UTC)
The previous discussion was about the prototype. This section begins the proper discussion of this template
Block design (algebraic combinatorics) is of mainly specialist interest, so the "(algebraic combinatorics)" serves as a warning (to the non-mathematical reader) and service (directing the reader to the most relevant articles with more information).
The article on "algebraic combinatorics" has little relevance to block designs. Is this link compatible with Wikipedia guidelines? 16:11, 23 June 2009 (UTC)Kiefer.Wolfowitz (talk) 23:24, 23 June 2009 (UTC)
My thought it that this has become far too cluttered with things pushing out into more general statistics and should be resticted to things that are directly "experimental design" topics. I suggest removing the groups of articles:
- null & alternative hypothesis · type I & II error · significance · p-value ·
- deduction, induction & abduction · logic · falsifiability
- minimum-variance unbiased estimator · Gauss-Markov theorem · least squares · maximum likelihood estimation
- F-distribution & test · chi-square distribution · Cochran's theorem
etc. It may be worth including a few articles on the major model types, but extending the template to include inference seems to go too far. There are other things that may also be making the coverage too general. Melcombe (talk) 14:54, 25 June 2009 (UTC)
- Thanks for your comments. It does seem too cluttered to me also. Below, I defend a few items, and for the other items explain why I included them. (I create subcategories for discussion: I hope this is good WP practice, and apologize if this is nonstandard.)
null & alternative hypothesis · type I & II error · significance · p-value
Type I & II error needs to stay, because reducing the probability of a Type II error is the usual concern for minimizing the number of replications (for ethical & economical reasons).
deduction, induction & abduction and falsifiability
There are a lot of uses of experimentation besides hypothesis-testing or estimation, and Box and others stress the importance of experimentation as a way of generating ideas, etc. However, I agree that these topics are inessential and contribute clutter.
Statistical criteria & methods: minimum-variance unbiased estimator · Gauss-Markov theorem · least squares · maximum likelihood estimation
Experiments are designed to satisfy statistical criteria, which are worth listing. The criteria & methods for estimating fixed-effects models deserve listing (imho) on their own-merits, and because their extensions for random effects follow on the next line: BLUP, REML, Bayesian hierarchical models. These random-effects topics are essential for repeated measurements experiments, and the public (and even statisticians) need some guidance about the principles (since these principles disagree for random-effects, unlike most linear fixed-effects models).
ANOVA "F-distribution & test · chi-square distribution · Cochran's theorem"
These seem to me essential for analyzing data from experiments. I would rather eliminate the MANOVA (multivariate)-related entries, which are most useful for graduate-students in statistics, psychology, etc. The t-test can be eliminated (next to Hotelling), but that would save only a few characters. Kiefer.Wolfowitz (talk) 15:23, 25 June 2009 (UTC)
Create separate Template for Linear Models and Statistical Inference?
Perhaps we could remove some entries in "models and inference", after creating another Statistics Template, say "Linear Models and Statistical Inference"? I would object to removing all of the inference topics, because inferential principles guide the design. Further, the design should guide the analysis (as we were taught by the books of David Freedman and Oscar Kempthorne!). Zealously, Kiefer.Wolfowitz (talk) 16:00, 25 June 2009 (UTC)
Example of a template with fewer entries
I believe that this template has about the same complexity as the other templates in our WP:Statistics project. (Orthogonal arrays are important, but the article is weak, and accessible through e.g. Taguchi methods). Kiefer.Wolfowitz (talk) 16:24, 25 June 2009 (UTC)
- I commented-out (hid) roughly one-third of the entries, reflecting the good-sense of Melcombe's suggestions. I trust that most of these were rather obvious, in retropsect. Thanks. Kiefer.Wolfowitz (talk) 17:16, 25 June 2009 (UTC)
- At this stage I would suggest moving to the "Navbox subgroup" type of structure that is found in the statistics navbox template. One could then aim to have at least two articles in a subgroup ... otherwise the subgroup shouldn't be a subgroup. One article I would question is subsampling, as this is just a disambiguation page and you would prefer to have something more direct. Melcombe (talk) 09:18, 26 June 2009 (UTC)
- Again, I'll move to the Navbox subgroup style soon. Some users have contributed suggestions about the content, which I'd like to fix first, e.g., like "subsampling versus replication".
- The embolding highlights the lead concept, which introduces the topic of the entries that follow it. I believe that this will be less useful with the semantical formatting of the Navbox sugroup structures. In the middle of the line, the emboldinging highlights an especially important topic, often independent of the preceding entries; some of the mid-line emboldened topics are followed by related specialty articles: Thus the embolding serves (in the revised Template) as compact way of presenting two semantic groups on one-line. (In most cases, there are additional entries which are rather advanced, which are hidden in the template now; perhaps they would be restored if we move to the semantic-formatting of NavBox.) Thank you.Kiefer.Wolfowitz (talk) 08:54, 29 June 2009 (UTC)
I'll explain why I linked subsampling to a disambiguation page. SubSampling is a very important technique to reduce observational error, and is often confused with independent replication. It deserves its own article (brief), and both of the disambiguation articles contains something relevant. Maybe including subsampling in the glossary is a better solution?Kiefer.Wolfowitz (talk) 13:28, 27 June 2009 (UTC)
- The entry now reads Replication versus subsampling. It would be great for somebody to consider a brief, focused edit on the replication article, to help readers from the Experimental Design template. Kiefer.Wolfowitz (talk) 19:52, 27 June 2009 (UTC)
- The prototype's style guided the current version. Perhaps you are right about the capitalization. Thanks for the suggestion. Kiefer.Wolfowitz (talk) 10:44, 29 June 2009 (UTC)
Redirect of "experimental design" to this template
I am troubled by the redirection of experimental design to this template because it is a realistic search term, as well as a realistic link term, although everyone should check where links go before saving them to an article. I suggest that "experimental design" be redirected to design of experiments and that links that currently exist to "experimental design" be changed, where necessary. There are currently 66 links, so it would be nice if the work were split among three to six people. -- Kjkolb (talk) 07:41, 5 July 2009 (UTC)
- I am sorry if this redirect was inappropriate. On the other hand, design of experiments is on top of the Template, so that any inconvenience should be minimal for those readers best served by the design of experiments article. Is there any way to observe how many searches go from this page to Design of experiments versus another linked article (judged to be more relevant)?
- I could undoing my redirect later this week (Thursday or so), if agreeable, and if it's simple to do. (Maybe I could do a first redirect on Thursday, and then the rest on Friday if no disaster had occurred.)Kiefer.Wolfowitz (talk) 10:13, 5 July 2009 (UTC)
Edits by Melcombe
I agree with most of the edits, and applaud the cleaning-up of hidden topics, which should make it easier for future editors.
On the other hand, I am concerned about the removal of
- Principles of the analysis of variance and regression, e.g. Gauss-Markov, BLUE, and Maximum Likelihood. I don't understand why Least squares is left, when it is a justified by the Gauss Markov theorem and BLUE.
- The criteria/methods of BLUP and REML are needed for mixed models and repeated measurements.
- The (non-central) F-distribution and F-tests are very important to the analysis of variance and to the planning of experiments, being necessary to planning the number of repicates to achieve power.
Also, let me suggest some principles for organizing the categories
- Kempthorne's: distinguishing blocking ("error control"), treatment, and observational structures (and block-treatment interaction with generalized randomized block designs, GRBDs) under the hypothesis of unit-treatment additivity.
- Scales for treatments, and possibly for covariates or responses: nominal (anova) and numeric (response surfaces and regression).
- I guess "least squares" got left because I didn't spot it. The initial intention was to leave only things immediately associated with design of experiments to see if there was some real structure present, but I don't really see one. A major problem is that the design of experiment article(s) don't really provide much structure to the overall topic. For the template, it would be good to find some outer-grouping names that would be reasonably understandable to the general reader, possibly: "design to detect effects", "design for best output", "size of experiments". I suggest leaving putting in the "fine detail" such as "F-tests" untila reasonable outer structure for the template can be found. Melcombe (talk) 10:04, 22 February 2010 (UTC)