Messages

Thanks for your efforts with the four-hour exam part, June 2. The Exam Project is now on the course website, with deadline Mon June 19, 11:59, to the reception room at the Department of Mathematics, in duplicate, please.

Note that each report should include two special pages (cf. the text, page 1). *Page A* is the self-declaration form (basically saying that yes, I've done this myself, without cooperation with others, and yes, when I've found material somewhere, I give references). *Page B* is a one-page summary of how you've worked with the project (what you found complicated, or easy, or inspiring, or challenging), and also a brief self-assessment of its quality.

Good luck with your work.

Nils Lid Hjort

On Tue May 16 I first spoke a bit on the (perhaps surprisingly dominant) role of model selection issues and the grander discussions of the Scientific Committee of the International Whaling Commission in Bled, Slovenia, May 10-14, where Celine Cunen and I took part, Moby-FIC-ing away. I then went through Nils Exercises 15 and 19, and discussed issues of Ch 7.

As I also explained, I & all other model selectors at the 8th floor will be busy with FocuStat's Building Bridges at Bislett workshop, Mon-Tue-Wed May 22-23-24:

http://www.mn.uio.no/math/english/research/projects/focustat/news-and-events/workshop_program17_web.pdf

But on *Fri May 26 from 14:00* I will teach for about an hour, in B81, and we go through the four-hour exam of stk 4160/9160 from 2015, now placed at the present course website. Note, incidentally, that you're allowed to bring along *one page of hand-written notes* for the four-hour exam June 2.

On Tue May 9 I went through Exercises 13 and 14 in Nils collection, and roughly the first half of Chapter 7. Next week I will round off Chapter 7 (where the last two sections are not part of the curriculum).

Exercises: 15 and 19 from Nils collection.

Greetings from Bled, Slovenia, where Celine Cunen has reported to the Scientific Committe of the International Whaling Commission (and delegates from a great many countries) about FIC for Whales (and where model selection issues have been crucially important and widely debated):

http://www.mn.uio.no/math/english/research/projects/focustat/publications_2/cunen_walloe_hjort1.pdf

https://www.facebook.com/groups/1589206911336271/permalink/1891603601096599/

Last week, Gudmund Hermansen took you through Exam 2009, Exercise 3.

For Tue May 9, try to go through Exercises 13 and 14 in the Nils Collection of exercises. I will also go through the first half of Chapter 7, on model averaging etc., the last "real" chapter in the curriculum.

I'm at the BFF conference (Best Friends Forever: Bayes, Frequentist, Fiducial) at Harvard May 0-1-2-3, but my PostDoc Gudmund Hermansen will teach *approximately one hour*, 9:15 to 10:15, on Tue May 2. He will go through Exam stk 4160 2009, Exercise 3 (how attendance data are influenced boy girl-boy, by math score, by art score; and how such data can be modelled, outside the plain Poisson).

1. On Tue April 4, I used most of the time for Exam 2011, the South African men and their hearts and their LDL scores. I also discussed various FIC issues more generally.

2. I've uploaded com16b and com16c, pertaining to the heart of the matter for these exercises; check them, run than, understand them, modify them.

3. There is no teaching in the Easter week, and also no teaching on Tue April 18. For Tue April 25, work through Exam  2011, Exercise 1, Exam 2013, Exercise 1. I will round off the FIC chapter 6, with a bit of attention to the AFIC.

4. What remains for May is then Chapter 7, on model averaging, plus "knowledge consolidation" and repetisjon, where we also do more exercises, on AIC, BIC, FIC, AFIC, each of which will show their flags & faces for the Exam Project in June 2017.

1. On Tue Mar 28 I went through the essentials of FIC, from Chapter 6, with more to come next week.

2. For Tue April 4, work with the following. From Exam stk 4160 2011, Exercise 2 (but not yet (f)); Exercise 3; then, if time permits, also Exercise 1.

1. On Tue Mar 14 I went through the core part of Ch 5, and did exercises, including a bit more about the mammals. I computed p1 = Pr(our brains are as big as they are | our body weight), for models 1 and 5; and p2 = Pr(water opposum have brains as small as they are | body weight), again for models 1 and 5.

2. I'll upload a "com14f" R script about this.

3. Next week I'm in Bruxelles (to help decide who shall have millions of Euro and who shall not). Celine Cunen teaches in loco parentis. She will teach from the start of Ch 6, the FIC chapter.

4. In addition, she'll go through decent parts of Exam STK 4160, 2015, Exercise 1 (a)-(e) and Exercise 3 (a)-(e). Work through these in advance.

1. On Tue Mar 7 I spent time going through various models and details for the mammals dataset, including assessments of degree of outlyingness for human sapiens (our brains are too big). I also presented the basic Start Theorem of Ch 5, with more to come next week.

2. I've placed com14b and com14e on the website, concerning the n = 62 mammals and the four or five models. Run through them, make sure you understand what goes on, so that you can "copy and modify" when needed.

3. Exercises for Tue Mar 14: First, fit Model 5 to the mammals dataset, with linear mean beta0 + beta1 x and varying \sigma_i = \sigma \exp(\gamma_1 w_i + \gamma_2 w_i^2), where w_i = (x_i - xbar)/sx. Draw a 99% confidence band for E(Y | x), and translate back to (x0, y0) scale, i.e. body weight (in kg) and brain weight (in g). Estimate p = Pr(Y_0 at least 1320 g | x0 = 62 kg), which is a measure of our outlyingness. Do the same with the water opposum (with its small brain).

Then a...

1. On Tue Feb 28 I went through Sections 4.1-4.2, and also various aspects of Poisson regression, the sandwich matrix for regression, overdisperson, etc. We did the birds on islands exercise #20.

2. I've uploaded "com13a" for the Poisson models for birds on islands.

3. For next week, we're doing (body weight, brain weight) for n = 62 mammals. Write

library(MASS)
x0 <- mammals[ ,1] # body  in kg, range 0.005 to 6654.000
y0 <- mammals[ ,2] # brain in g,  range 0.14  to 5712.00

You'll find your own modest person as no. 32 in this list. Transform to (x,y) = (log x0, log y0), and then carry out model fitting and selection for y vs. x. Include a total of four models: (i) ordinary y = beta0 + beta1 x + N(0,\sigma^2); (ii) quadratic regression y = \beta_0 + \beta_1 x + \beta_2 x^2 + N(0,\sigma^2); (iii) linear regression but with varying \sigma, with \sigma_i = \sigma exp(\gamma (x_i - \bar x)); (iv) the widest model, with both...

1. On Tue Feb 21 I rounded off Ch 2, with cross validation and other comments, before I gave a mini-version of Ch 3. I gave the BIC derivation and briefly discussed (i) Landsteiner 1924 blood groups in man and (ii) my analysis of Solzhenitsyn vs. Sholokhov. The full paper about this is here:

cas.oslo.no/getfile.php/138668/CAS_publications_events/CAS_publications/Seminar_booklets/PDF/Consilience_LidHjort.pdf

2. I also discussed the babies & mothers dataset, with model selection among 2^3 = 8 logistic regression models, etc.; see the R script com12b.

3. For Tue Feb 28 do the following.

(i) Make sure you understand and can use com12b and moficiations. Check what happens with a different mother than Mrs Jones (create her yourself). Supplement the AIC analysis with AIC^*, for which you need \hat p^* = Tr(\hat J^{-1} \hat K) for each of the 8 models. Derive the relevant formulae for \hat J and \hat K (see Ch 2), and find the \hat p^* numbers....

1. On Tue Feb 14 I went somewhat briefly through the basics of regression models and the AIC and AIC^*. We'll be coming back to parts of this in connection with exercises. This is partly in the Aristoteles spirit: "For the things we have to learn before we can do them, we lean by doing them." I also gave a mini-summary of BIC, with more to come next week. We also spent time on the nerve data exam question from 2015, and other details.

2. Exercises for Tue Feb 21: Go to Claeskens and Hjort book website

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

and download the dataset on small babies; check the description of the various components. Concentrate on the variables x = weight of mother in kg, z1 = smoke (1 for smokers, 0 for non-smokers), z2 = ftv (number of visits to physician during the first trimester), z3 = indicator for "race" being "white", which you get by writing z3 = 1*(race==1). With y = indica...

The exam for this course consists of two parts:

Part I, The Project: tentatively scheduled to have t0 = Wed June 7 = the date where the exam set set is made available, and t1 = Mon June 19 = the date where candidates need to deliver their reports.

Part II, 4 hours written exam: Fri June 2.

Please report immediately to Nils if the (t0, t1) time window above is seriously bothering you.

1. On Tue Feb 7 I went through (i) the delta method, for inference about a focus parameter mu = mu(theta); (ii) the basic AIC and AIC^*. I also discussed other material from Ch 2, and discussed two ways of mixing up the Weibull -- on the f scale and the cumulative hazard scale, with application to the Old Egypt dataset.

2. I'm uploading R scripts com3c (old Egypt, mixing Weibulls); com5a (polynomial regression, Exercise 2); com6a (old Egypt, gamma-weibull-extended model). Run them, check that you understand their components, so that you may modify these scripts for later purposes.

3. Next week: I round off Ch 2, with regression models. Exercises: Do #2 (and in that connection also relevant bits of #1). And do Exam Project 2015, Exercise 2, a-b-c, involving the nerve data with n = 799 datapoints; this dataset is found on the 2015 version of the course website.

1. On Tue Jan 31 I worked my way through certain important parts of Ch 2, and also discussed aspects of modelling and logLik-maximisation for the Egypt life-lengths dataset. Three things will be discussed next week, to complete the picture: (i) How to turn \hat\theta information and precision into \hat\mu information and precision, where \mu = \mu(\theta) is a function of the model parameter vector. This involves the *delta method*. (ii) Extending theory and methods from iid to regression models. (iii) The necessary theory behind the AIC and the model-robust AIC.

2. I've placed "com3c" on the website, doing 0.5 weib + 0.5 weib mixture for the Egypt dataset. Try it out, and do some modifications. Make sure you understand how the details work.

3. Exercises for Tue Feb 7:

(i) Make one more attempt at modelling the Egypt dataset, with n = 141 iid observations. Again, try to find models of parameter size 3 or 4, which beat the best of Gamma and Weibull. Include...

1. On Tue Jan 24 I gave a broad introduction to the course and to the curriculum book, "Model Selection and Model Averaging", by Claeskens and Hjort, and which you need to get hold of. I started on Ch 2, with ML theory and the AIC formula, and will continue this path next week.

2. I've placed the R script "com2a" on the course page, along with the Ancient Egypt lifelengths dataset. Make sure you can run the script (via writing source ("com2a") in R) and that you understand all details of what goes on. (There was a minor error in an earlier version, so make sure also that you're running the corrected code.)

3. Exercises for Tue Jan 31: First, using extended versions of "com2a", try out a couple of other models for the Egypt data, with n = 141 lifelengths (and so far we do not split these data into men and women groups). Try in particular to come up with a three-parameter model which beats the Gamma and the Weibull, in...