Messages

The oral examinations Tue June 14 take place in seminar room B70, with time schedule etc. as indicated in yesterday's message.

Thanks for your efforts (so far)! We shall quite enjoy reading your reports over the weekend. Oral examinations are arranged Tue June 14 as follows (a later message will give precise information concerning which seminar room is to be used):

09:30 Martin Lerujordet

10:10 Martin Jullum

10:50 Peter N. Maisha

11:30 Marianne R?ine

12:30 Rune Hoff

13:10 Lars Sevald Leknes

Note: There is an unfortunate inaccuracy in Exercise 1(h) (stemming from me first working with a different focus parameter than the one I chose to use in the end). The focus parameter mu is meant to be equal to the upper tail probability Pr(Y \ge y0) throughout Exercise 1, as explained in (b)-(g), i.e. also in point (h). So \mun there is as in the formula of point (e), but with a = delta/rootn, etc. -- Though this is fairly clear from the context and hopefully does not stop anyone from thinking along the right lines, please mention this correction detail to your comrades.

[Just to avoid any potential confusion: I first put up a version of the Exam Project at 11:30, but took it out to repair a minor detail before putting it back in at 11:45. Make sure you have the final version, which is the one available now.]

I'm now making the Exam Project available here, and wish all candidates a suitably fruitful and not too time-consuming struggle. Deliver two copies of your report to the administration, by Thu June 9 at 14:00 at the latest. Oral examinations will be arranged Tue June 14; more information about this will come later.

The data-file for this exam project is also available here ("ldl-data"); you need to download this to your computer.

Note that you are required to include two extra special pages with your reports (see page 1 of the project document) -- Page A (the usual form, declaring that this is your own work and that you have been working alone, etc.), and Page B (your one-page summary of your efforts, including self-assessment of the quality of your report).

If in doubt about anything, send a mail to NLH. The candidates are also advised to have a look at these pages in case there is...

I'm placing the R script "com107b" here, for Exam 2009 #3. It does Poisson regression plus a couple of other relevant models, and gives aic, bic, tic, fic, afic. Note that the pstar values are big for the Poisson regression models, reflecting overdispersion (but the two last models do much better in this regard).

For some reason the usual editing tools do not work today, so I'm not yet able to organise a separate little section for com files; it's so far placed under "Examination".

Note also a suitable list of R com files under the 2009 course website.

Exercises for May 10th (one hour lecture + two hours exercises): We complete those listed below for May 3rd, and also work through Exam 2009 #3 (a)-(b)-(c).

Exercises for May 3rd: For the n = 799 nerve impulse data of Exercise 7 and 10, consider the four focus parameters of 10(c). Compute estimates and FIC scores, for the four candidate models, using the exponential 1-parameter model as narrow model and the three-parameter as the wide one. Also display a FIC plot. -- Also, do Exercises 17 and 18.

Exercises for Gagarin Day: Exercise 10/2009, and the following: Consider the version of the log-linear expansion model of Exam 2009 which uses only one additional parameter aj, i.e. for one extra psij term. Compute the tolerance radius, for each of j = 1, 2, 3, ..., 10.

Important: Tentative time schedule for the examination process is as follows. (1) Project exam, made available Mon 30 May, reports to be handed in Thu June 9. (2) Oral examination, ca. 35 minutes per candidate, on Tue June 14.

Please contact Nils LH immediately if the above schedule may be difficult for any of you. By next week we finalise this.

We have started Ch 5.

For the coming two weeks we shall go through the following exercises. (i) The remaining details regarding Exercise 9/2009. Estimate the 0.10-quantile using the two models, along with confidence intervals. (ii) Let y1, ..., yn be iid N(0, sigma^2), and consider models M0: sigma = sigma0 = 1 and M1: sigma is positive. Work out precise aic and bic recipes, and try to emulate the two central dimensions of the "prototype story" from Ch 4. Work with Pn(sigma) = Pr(selecting model M1 | sigma) and with risk functions for estimating sigma, and display curves, using e.g. n = 100 and n = 1000. (iii) Exam set 2009, 1(a)-(g) and 2(a)-(c).

I rounded off Ch 3 and started Ch 4, where only 4.1-4.2 are inside the curriculum. Next week I plan to finish these sections and start Ch 5.

For March 22, do Exercise 9/2009 and also Exercises 2 and 3(a)-b) from the 2009 Exam Project.

We've nearly concluded Ch 3 and will start Ch 4 next week, where only Sections 4.1-4.2 are inside the curriculum.

For March 15, do Exercise 9/2009, where you also compute the BIC score and associated posterior model probabilities; conclude Examples 3.6 by also including the 3rd model that simply keeps thetaA, thetaB, thetaAB, thetaO as unknown parameters; and finally set up a framework for comparing survival analysis models of the type hi(t) = exp(beta0 + xi'beta), complete with an implementation that gives ML analysis along with AIC and BIC scores. As test data set for your programmes, try the "primary biliary cirrhosis" one from the book's website.

We have started Ch 3, where Sections 3.1-3.4 are curriculum, not 3.5-3.6. I plan to move relatively quickly on to Ch 4, where only Sections 4.1-4.2 are curriculum, and then to Ch 5.

For March 8 we do Exercise 9/2009; Exercise 3.5; Example 3.6 (and try to compute the exact posterior model probabilities, starting from prior model probabilities 0.5 and 0.5 and with uniform priors for (a,b) and (p,q), respectively); and, finally, the following:

Let y1 and y2 be independent from Bin(m1,theta1) and Bin(m2,theta2), and entertain two models: M1 says theta1 = theta2, M2 says otherwise. Take prior probabilities 0.5 and 0.5 and flat uniform priors for the model parameters. Give formulae for Pr(M1 | data) and Pr(M2 | data), and compare with the BIC formulae. Illustrate with (a) y1 = 7 and y2 = 4, with m1 = m2 = 10; and (b) y1 = 58 and y2 = 42, with m1 = m2 = 100.

I'm in Arizona the week of 14-20 Feb, and there is no lecture Feb 15.

For Feb 22 we go through the AIC and relatives parts of Ch 2. Also, prepare exercises 6, 7, 8 from the 2009 collection -- with the extra challenge for you to invent your own alternative model for the 799 nerve impulse data.

We've started Ch 2, and key wors so far include maximum likelihood, Kullback-Leibler, information matrix, least false, LLN (law of large numbers), CLT (central limit theorem), delta method. We also went through Exercise #4/2009, and I'll place the R programme "com2f" on the webpage. Make sure you understand all details here, as you'll be required to go through similar adventures on your own later on.

For Tue Feb 8, supplement this Exercise 4 with standard errors and confidence intervals computed without assuming that the models are correct (cf. the J^(-1)KJ^(-1) quantities). Then do Exercise 5, but with focus parameter p = Pr(X \ge 0.222).

Today I went through some of the more general issues touched on in Ch 1, and just had time to start Ch 2, where we continue next week. We also went through Exercise 2/2009.

For Tue Feb 1, do Exercise 4/2009.

1. Today I gave a general but brief introduction to the course, including information about various practicalities.

2. For Tue Jan 25 I plan to use 1 hr on Ch 1, 1 hr to start Ch 2, and 1 hr for us to go through Exercise 2 from the 2009 collection. This exercise requires results and details from Exercise 1, which you may also work through, of course, but that Exercise will not be dealt with in detail here.

You're all warmly invited to the stk 4160 course on statistical model selection. We start Tue Jan 18, 9-12, in seminar room B 81. You need the course book "Model Selection and Model Averaging" (Gerda Claeskens and Nils Lid Hjort, Cambridge University Press, 2008), which may be found in the Akademika bookstore.

You are also advised to browse through the Spring 2009 version of this stk 4160 course site. In particular, print out the "Exercises and Lecture Notes" there (version E, 32 pages), for your convenience, and it will also be useful to check out the eleven-day exam project from that year.

Further practical information will be provided on Wednesday.