Messages

Again, good luck & good work, everyone -- you find the pdf for the Exam Project on the course webpage, and you also need to access & work with two datasets, sorted under "some datasets". These are called "sweden-accidents" and "fivektenk-data".

Just a little footnote -- on various previous occasions, in many Master- and PhD-level courses, I've written about "Page A" (self declaration form, that you haven't been cheating, etc.) and "Page B" (the one-page summary of your work, how it has proceeded, what has been easy and what has been difficult, and how satisfied you are with the quality of the report).

The system parameters are a bit different for this occasion, since students deliver their reports qua a pdf (as in one and only one pdf) to the Inspera system. Hence I'm goring to rephrase a little. (a) There's no "Page A", but something on the last page of my Nils Project (?), saying that you delivering your report is taken to mean that you've been working alone, haven't copied from others, been working on your own, etc. (b) So there's no "Page B", but something equivalent:

" Importantly, by handing in your report to the Inspera system you guarantee th...

So the last round of teaching is Friday May 31, 14:01 to 16:01, in our usual lecture room. We discuss bits & pieces of the curriculum, and I'll try to answer any questions & queries in real time.

The four-hour written exam is Tue June 4, 9-13; check the details regarding Silurveien etc. Then you should bring a simple calculator, enough to compute the square-root of 49 and some exp and some log, *and* one single sheet of paper, with your hand-written notes.

The Nils Project (insert TradeMark symbol here) is made available on this course website Wednesday June 5, and you need to produce & upload pdf versions of your reports by Monday June 17 at 14:29 (or earlier). This is to be done inside the Inspera system.

For the Nils Project, i.e. Exam, Part II, there are two mandatory pages you need to include: Page A is a signed version of a "selverkl?ring", which I'll upload to the course website in time, saying you've been working on your own, that you ha...

There's teaching as usual Wed May 29th, 13-16, and also two extra hours on Fri May 31st, 14:15 to 16:00, in our usual B 819. For these two last teaching events of the course we use time also for *summing up* and for questions-and-answers. So please use the occasion to think through the various parts of the course, and with questions. 

Exercises I plan to go through: First, we start with the *four-hour written exam 2015*, which is found on the course's website for 2017. Then, exercises 1, 3, 2, in approximately that order of priority, from the exam set of 2009.

I'm away Wednesday May 15, and it turns out my intended "vikar" can't make it after all -- hence there's no teaching that day.

For Wed May 22, do the exercises pointed to in the previous message, and then I'll also go through the essence of what remains from Ch 7. Key words there include the Quiet Scandal of Statistics and behaviour of post-selection and model-average estimators.

There's also teaching Wed May 29, and we find there's a need for it we can find two extra-hours of going through exercises, before the exam period sets in.

1. P? seiersdagen 8.5, C¨¦line Cunen went through details for Exam 2011, Exercise 3 (South African men, their LDL, three-box probabilities, various models) and Exam 2015, Exercise 3 (Danish melanoma, hazard rate models).

2. R-code is available, in com20a (Nils) and com202a (C¨¦line).

3. For Wed May 15, do Exam 2009, Exercise 3 (attendance data, with scores, Poisson, etc.), and Exam 2017, Exercise 2 (long lives in Scandinavia). More details in a little while.

1. On Wed April 24 I discussed various matters on the broad boarderline between Chs 6 and 7, so to speak, showing various applications following from the master theorem and \rootn (\hat\mu_S - \mu_\true) tending to \Lambda_0 + \omega^\tr (\delta - G_S D). I also talked through aspects of Exam 2011 Exercise 2, LDL data for South African men.

2. On the slava trudu day there's no teaching -- but honouring work is allowed to involve honouring work.

3. On Wed May 8 I'm in Bruxelles attempting to hand out millions of Euro. C¨¦line Cunen will teach in loco parentis. She will tell you bits & pieces from Ch 6, but will focus on these exercises from earlier exam projects:

(a) 2011, 3 (more about the South African men and their LDL); 

(b) 2015, 3 (Danish melanoma, hazard rates);

(c) if time, 2009, 3 (attendance in school, Poisson type regression).

1. On Wed April 3 we did the essential parts of Ch 6, with FICology. We also went through Exam stk 4160 from 2011, Exercise 1.

2. As agreed then, for Wed April 10 we go to South Africa, to study the LDL levels of a population of higher-risk men, i.e. Exercises 2 and 3 from the exam project 2011. Do as many of the points as you can.

1. We've started Ch 6, with the last part of the course, and the master theorem is about the limiting normal distributions for each submodel setimator \hat\mu_S, with biases and variances. This leads to the FIC. We also did Exam 2013, Exercises 1 and 3. 

2. We'll be FICin'.

3. For Wed April 3, do the risk functions for Exam 2013 Exercise 1 more accurately, and supply these with that for the post-AIC estimator. Then do Exam 2015, as much as you can for Exercise 1, and Exam 2011, as much as you can for Exercise 2.

4. Check Jens Krisoffer Haug's talk at PRIO, a few weeks ago, about FIC applied to armed conflicts data models, and check how he explains the difference between AIC and FIC:
https://www.mn.uio.no/math/english/research/projects/focustat/talks/

1. On Wed Mar 20 I went, somewhat quickly, through the last parts of Chapter 5, and tended to Nils exercises 7 and 9. The last part of the course corresponds to Chapters 6 and 7, and we start with Chapter 6, FICology de Lux, next week.

2. Exercises: find the Exam set for 2013, and work through as much as possible of Exercises 1 and 3. To illustrate part of the story of Exercise 3, simulate n = 400 points from the model f(y, \theta, \gamma), with \theta = 3.333 and \gamma = 1 + 1.444/\sqrt{n}. Estimate the median of the distribution, under the narrow and the wide model, along with confidence intervals. For this median parameter, which of the three estimators worked with in the exercise appears to be the best?

1. On Wed Mar 13 we did various things for Chapter 5, tolerance levels and regions and ellipses and strips around home models, and with the master theorem describing how estimators using respectively the narrow and the wide model perform. We also did extra exercises, e.g. tolerance around the normal for a certain skewness, and tolerance for Bj?rnholt's skiings days in two different directions.

2. I've placed the R script com21a on the website, with numerical matters for tolerance around the normal with respect to the skewed normal I invented. It involves numerical integration in R.

3. Exercises for Wed Mar 20: Nils Exercises 9 and 7 (in approximately that order). For Exercise 9, do the theory parts first, and to create a dataset to work with, do the following: simulate 200 points from the normal mixture, 0.5*N(-0.5,1) + 0.5*N(0.5,1). Based on your tolerance level calculations etc., which of the two models would be be...

1. On Wed Mar 6 we discussed the first part of Ch 5, with tolerance intervals around parametric models, along with examples. I also went through parts of extra exercises, skiing days at Bj?rnholt plus exam 2015, exercise 3, the Danish melanoma study.

2. I've uploaded *com19b* (skiing days at Bj?rnholt, two models) and *com20a* (for the Danish melanoma data). Run them, read and modify, part by part, and do more.

3. For Wed Mar 13, do these extra exercises, related to tolerance regions around models.

(a) "Grow the normal" by including the extra parameter \gamma, in F(y,\xi,\sigma,\gamma) = \Phi( (y-\xi)/\sigma )^\gamma. Draw densities for \gamma equal to 0.9, 1.0, 1.1. Compute skewness as a function of \gamma. How much must \gamma differ from \gamma_0 = 1, in order for the three-parameter model to be better than then classic school-book normal model?

(b) Go skiing to Bj?rnholt. Take the home model to be cla...

1. Our time window for the Exam Project is [t_0, t_1], with t_0 = Wed June 5 and t_1 = Mon June 17.

2. On Wed Feb 27 I went through the essentials of Ch 4, via Sections 4.1-4.2, with AIC vs BIC. I also went through the exam 2015 exercise 2, on stretching parametric models, plus more of the football modelling exercise, etc.

3. I'm in the process of uploading R com scripts for a couple of recently discussed exercises.

4. For Wed Mar 6:

(a) Find the Exam 2015 set, and attempt to do as many points as you can for Exercise 3, the Danish melanoma data. There are a few "FIC points" which we need to postpone a few weeks, but you can do most of the points.

(b) Go to the course webpage for STK 9190, Spring 2018, and the exam project's exercise 2, with (calendar year, number of skiing days at Bj?rnholt). Model this dataset with so far two models: (i) ordinarly linear regression, y_i = a + b x_i + \sigma \eps_i, with the \eps_i being i.i.d. standard normal;...

1. We need to land on t_0 and t_1 for the exam project time window. The "first iteration" is t_0 = Wed June 5 and t_1 = Mon June 17. We converge to consensus next week.

2. On Wed Feb 20 I rounded off Chapter 3, with a bit of time for some of the application stories (who wrote §´§Ú§ç§Ú§Û §¥§à§ß, what's the memory length of Ja vi elsker dette landet?), and went through several exercises, including the log-polynomial modelling for mhaq data, from Exam 2017 Exercise 1. 

3. On Wed Feb 27 I plan to go through sections 4.1-4.2, the core story of Chapter 4. After that we're in for the last third of the course, namely Chapters 5, 6, 7, about tolerance regions around models, about FIC, and about model averaging.

4. For Wed Feb 27 attempt models for the football data, with n = 254 matches, giving (y1, y2, fifa1, fifa2). Take the two-parameter model with Poisson rates \lambda_0 \exp(\beta fifaratio) as "t...

1. On Wed Feb 13 I rounded off Ch 2 and did the essential BIC approximation thing for Ch 3. We did J and K analysis for the normal, learning both about normal approximations and about which standard results for the normal that cannot be trusted if the model is not correct. I also spent time doing extra exercises (a) KL(g, f_\theta) for given g, with various illustrations, and (b) modelling football results, via (y1, y2, fifa1, fifa2). For these, check com15b and the file football_data on the course site.

2. Next week I more or less round off Ch 3.

3. Exercises for Wed Feb 20 are as follows.

(a) Find the Exam Project 2017 and do Exercise 1.

(b) Work with the 254 football matches, with data to be organised into (y1, y2, fifa1, fifa2) (with data collected and organised by Nils in 2007). Try out some models, and don't give up until you have something working better, in terms of AIC, than the already satisfactory model which takes independent Poisson with...

1. I have now more or less finished Ch 2 (i.e. the curriculum active parts of that chapter), with AIC, AIC^*, cross-validation, regression models, Jhat, Khat, sandwich estimator, etc., but will spend a bit more time on Wed Feb 13 to sum up things.

2. I've placed com12a on the course website, with code doing the essence of Nils Exercise 2.

3. Exercises for Wed Feb 13 are as follows.

(a) Simulate say 1000 versions for Exercise 2, to track the probabilities for the different models winning the AIC competition.

(b) For Exercise 2, work out how not only the Jhat matrix can be computed, but also the Khat matrix, and hence the sandwich matrix.

(c) Suppose the true density for a certain lifetime distribution is g(y) = 0.5 dgamma(y,a1,b1) + 0.5 dgamma(y,a2,b2), with (a1,b1) equal to (1.238,0.061) and (a2,b2) equal to (4.504,0.111). For the two-parameter candidate models (i) gamma, (ii) Weibull, (iii) Gompertz, work out the least false parameters, and als...

1. On Wed Jan 23 I went through Nils Exercise 4, with Old Egypt, and more, and the basics of "the first half of Ch 2", with ML, how to find it, properties, the Law of Large Numbers, the Central Limit Theorem, how these may be used along with Taylor expansion and a bit more to demonstrate the crucial properties of ML estimators. This will be used to build the AIC etc. in weeks to come.

2. I've placed "com11a" on the website and will also upload "com11b", doing more models for Old Egypt.

3. For Wed Jan 30, do Old Egypt one more time, with the particular model that takes f = 0.50*gamma(y,a1,b1) + 0.5*gamma(y,a2,b2). Find the ML estimates, with their standard errors; draw two estimated densities on top of the histogram, namelig the "a single gamma" and the "equal mixture of two gammas"; compute the log-likelihood maxima; compute AIC scores; comment on your findings. -- Also, do Nils Exercise 2.

1. On Wed Jan 16 I gave a broad general introduction to the course. The curriculum is essentially Ch 1 (but "read only"), Ch 2, Ch 3, Ch 4 (but only 4.1, 4.2), Ch 5, Ch 6 (not all of it), Ch 7 (the first four sections), Ch 9 (mostly "read through").

2. You all need to get hold of the Claeskens and Hjort (2008) book.

3. I've placed Nils Lecture Notes & Exercises collection at the course website; print it out.

4. I've placed a simple R script com11a on the website; try it, run it, understand it, change small things and see what happens, copy and modify.

5. For Wed Jan 23, do Nils Exercises 4, 2. Use the com11a for the gamma(a,b) model, and amend and modify etc. to make programmes for the other models. For each, do simple analysis for the F^(-1)(0.50) and F^(-1)(0.80) quantiles, and compute logLmax and AIC values.