Messages

As previously announced & explained, this course will have a two-part exam. Part I, "The Project", will be made avaiable via this site on Wed June 3, before noon. You need to hand in your project report by Mon June 15, at 13:00, or any time earlier, to the Mathematics department's administration office at the 7th floor, *in duplicate*. 

Part II is a four-hour exam, taking place at Thu June 11 (see separate information regarding the precise time and location). This is a no-book exam, but you are allowed to bring with you *one sheet of paper*, hand-written or machine-made, with your own notes (only one side of the paper, not both).

Please do contact Nils if you encounter any practical problems with any of this.

I've uploaded

* com28f, for fic and afic analysis of the South African men, from exam 2011 exercise 2;

* com30b, com30c, com30d, for risk functions risk(delta) etc., both for limit calculations and "sanity check" for finite n, in a setting from exam 2013 exercise 3.

Thursday May 28 is our last teaching day. We use it for general discussion, summing up, pointing to certain crucial parts of the curriculum, etc.; please also feel free to ask questions etc. I will also give a bit of information regarding the exam (Part I & Part II).

We also do the last half of Exercise 3, exam 2013, where you *in addition* to the points given to the students then also should *test limit results against finite-sample performance*.

Specifically, for the limit risk functions for narr, wide, smooth, compute these curves for say n = 100 and n = 500, via simulations (at theta0 = 3.333, as per the text). To simulate data from the model, use the inverse cdf method. Also, simulate the distribution of the smoothed model averaging estimator \hat\mu^*, at a few places in the parameter space, both for n = 100 and n = 500, and using the limit distribution result from Chapter 7. 

In addition to summing up certain aspects of the theory of Chapters 6 and 7, we go through Exam 2011, Exercise 1, and Exam 2013, Exercise 3. For both of these exercises, add on an extra point, consisting in finding and displaying risk functions for finite n. The point is partly to check the quality of the large-sample approximations -- you may e.g. display risk functions for n = 100 and the large-sample limit in the same diagram.

We proceed with Chapter 7 material, including finer aspects of "the Quiet Scandal of Statistics".

Exercises: Work through Exam 2011, #2 and #3. After this we do Exam 2013, #3.

1. On Thu April 16 I went through various details related to Exam 2009 exercises 1 and 2, with FIC, tolerance regions, etc. I also summed up FIC and AFIC matters from Ch 6. Next week we start with Ch 7.

2. I have placed *com7e* and *com8a* on the site, related to Exam 2009 exercises 2 and 3. Note in particular that com7e is "generic" in nature, and suitably modified versions might work in general FIC settings, where 2^q submodels are worked through.

3. For Thu April 23, do Exam 2011 exercise 2, and if you have time, start also with the accompanying exercise 3.

1. This Thursday I will continue with Ch 6 methods, and in particular go from FIC to the AFIC.

2. Exercises: First Exam 2009, 3 c (where we earlier have been through points a-b-f-g). Then Exam 2013, 3 d-e-f (we have done a-b-c earlier). Finally Exam 2009, 1, as far as we get.

1. I will have a few more remarks concerning Ch 5 before embarking on Ch 6, the FIC chapter.

2. Exercises for this week: First, Exam stk 4160 for 2013, #3, a-b-c-d-e. Then from Nils collection, #10, as far as we manage.

1. Thanks to Martin Jullum for stepping in last week. He went through the first couple of sections of Ch 5 and Exercise 9.

2. This week I continue with Ch 5 material, with tolerance regions around parametric models, and go through Exercise 13 (as mentioned last week). If we find time we also start going through Exam 2013 Exercise 3, a-b-c-d.

I'm at a workshop on sakprosa on Thursday 12th (and have to talk about "the rhetoric of science", whatever that is). Two of my PhD students are certified model selectors and ficologists, however, so Martin Jullum and Sam-Erik Walker, in some combination, will teach this Thursday, from the start of Ch 5. For exercises, work with Exercises 9 and 13 from the Nils collection.

1. On Thu Feb 12 I discussed basic large-sample approximation methodology for regression models, also outside model conditions, leading to AIC, AIC^*, etc. We also went through more details regarding Exercise 4, the sandwich matrix, confidence ellipses, etc. Next week I intend to finish off Ch 2 and start with BIC in Ch 3.

2. Exercises for Thu Feb 19: First, a challenge: invent a three-parameter model for the 141 Egyptian lifelengths with a higher AIC score than the Gompertz (the "current winner"). Then, do Exercise 5, and 21 (a), about Poisson regression models. For the latter, find both AIC and AIC^* values, which requires finding both the J and the K matrices in question.

1. On Feb 5 I discussed ML large-sample theory, also outside model conditions, leading to the sandwich matrix J^{-1} K J^{-1}, and also derived the AIC and AIC^* formulae. Next week I proceed with more from Ch 2, including general regression models.

2. I have placed "com2a" on the website, with details pertaining to Exercise 4, fitting four models to Egyptian lifetimes.

3. Exercises for Thu Feb 12: First, do a bit more for Exercise 4, finding both AIC and AIC^* values, the latter using p^* values as opposed to merely p. Then do exercise 5 from the Nils Notes & Exercises.

1. On Jan 29 I discussed maximum likelihood, some large-sample theory, approximation theorems under model conditions, and spent time on Exercises 1 and 2. Next week I will go through large-sample ML theory outside model conditions and work with the AIC and its cousins.

2. Note that I have placed R scripts "com1a" and "com1b" on the website, for Exercise 2. Run them and check them and modify them as you might wish; in particular, run com1b with models allowed to be of order up to pmax = 20, to see what happens.

3. Exercises for next week: Please go back around 2100 years in time, to Roman era Egypt, observe life for men and women, and do Exercise 4.

1. On Jan 22 I gave a broad introduction to the themes & practicalities of the course. I sketched the core aspects of the book's Ch 1, and will start with Ch 2 next week, focusing on ML theory and applications, leading to the AIC.

2. Exercises you should work with for next week: #1, 2, from the Exercises & Lectures Notes collection.

Welcome, everyone, to the course on statistical model selection and model averaging. The course book is Claeskens and Hjort's "Model Selection and Model Averaging" (Cambridge University Press, 2008), which is or will be available at Akademika, or may be ordered from the publisher or amazon. During the course you will need to work with datasets from the book, available here:

http://feb.kuleuven.be/public/ndbaf45/modelselection/

Also, you may check the course website from the spring semester 2013, containing also a collection of Lecture Notes and Exercises. There are lectures 12:15 to 15:00 on Thursdays, starting Jan 22, and in most cases the teaching will be structured as two hours of lecturing plus one hour of working through exercises.

Nils Lid Hjort