Template talk:Least squares and regression analysis
|WikiProject Statistics||(Rated Template-class)|
As for "Experimental design" I think this has become too general. And yet it contains nothing about regression model/variable selection. I will remove some things that are neither regression analysis or least squares. At a general level, recall that this template will often occur alongside the statistics template. Melcombe (talk)
- We can delete Design of experiments, but I would like to keep Optimal design and Bayesian experimental design, which are especially about design for regression. However, those articles are rather technical, and most WP users would probably be better off with first viewing the Design of experiments article.
- Indeed, 1-3 months ago, there was discussion of revising the introduction to (or renaming?) the article Optimal design, to clarify that most of the article's description and probably most of the literature concern designs for linear or polynomial regression; even nonlinear regression models are approached often by linearization.
- The response surface methodology link should be kept (but I assume that RSM is approved by all). Kiefer.Wolfowitz (talk) 15:33, 25 June 2009 (UTC)
- I have added a "model exploration" group to allow regression model/variable selection to be covered and to include response surface methodology. I have also made Design of experiments a separate group (which I hope can be kept small). If the major groupings required can be got right it makes it easier to prune things down to reasonable coverage of articles ... at least it provides an indication of why they are there. Of the things I deleted previously, perhaps some could go in an "algorithms for implementation" group if it were sensible to have one. Melcombe (talk) 09:56, 26 June 2009 (UTC)
The line used to have these names:
- Foundations Statistical inference - Model selection - Minimum description length - Minimum message length - Algorithmic information theory
I think that since Kolmogorov suggested the information-complexity approach and since this approach has been useful for exploration and predictive inference (Rissanen, etc.), it would be better to restore MDL and MML (at least). It's better to follow Kolmogorov's suggestions than emulate the hoi polloi.
RSM belongs with experiments 
In general RSM is taught as a 2nd course in statistical experiments, and the textbooks are all written by specialists in design of experiments. Thus, I would argue that RSM belongs with experimental design (although Atkinson and Draper have written good books on variable selection and other regression problems). Kiefer.Wolfowitz (talk) 16:17, 26 June 2009 (UTC)
- I have moved some things between groups as suggested. I don't think Minimum description length, Minimum message length and Algorithmic information theory belong here but, if they do, they should be in a group whose name suggests why they might be of use in the context of "Least squares and regression analysis". They don't seem to be of direct practical use in this context and even a direcrt link in a theoretical context seems doubtful.Melcombe (talk) 17:08, 26 June 2009 (UTC)