FocuStat:

Focus Driven Statistical Inference with Complex Data CONTENT: FocuStat: The Task Force 2 The Tasks and Themes 3 The Task Focused Statistical Methods 4 Force Statistical Stories 12

Workshops and the VVV Conference 20

The FocuStat Research Kitchens 26

VIPs, Guests, Collaborators, Networking 28

FocuStat Excursions 30

The FocuStat Blog 32

Awards & Prizes 35

FocuStat in the Media 36 The core group, at a book discussion evening: G.H. Hermansen, E.Aa. Stoltenberg, V. Ko, Talks & Presentations 40 S.-E. Walker, B. Cuda, C. Cunen, K.H. Hellton, N.L. Hjort.

Publications 44

FocuStat after FocuStat 48 FOCUSTAT HAS BEEN a five- 2016, 2018; Ko, Stoltenberg year project within the Statistics and Walker will defend theirs Division of the Department of in 2019. After their PostDoc Mathematics, University of Oslo terms, Hellton is a senior (UiO), funded by the Research researcher with the Norwegian Council of Norway, operating Computing Centre, Herman- from January 2014 to December sen a First Amanuensis with 2018. The directly funded core the BI ­Norwegian Business group has consisted of Professor School; also, Cunen is from 2019 Nils Lid Hjort (project leader), onwards a PostDoc with the Kristoffer Herland Hellton and Department of Mathematics, Gudmund Horn Hermansen UiO. Former Master students (PostDocs), Céline Cunen and Josephina Argyrou, Jonas Moss, Sam-Erik Walker (PhDs). Riccardo Parviero, Leiv Tore Salte Rønneberg (again, super- SEVERAL ASSOCIATED PhD vised or co-supervised by Hjort, students and Master students, and each of them having taken along with further colleagues up PhD positions afterwards) and collaborators, have also have worked with FocuStat been close to the core group. In related themes, as have Martin particular, PhD students Martin Tveten and yet others. For a Jullum, Vinnie Ko, Emil Aas full list of associated members, Stoltenberg (each supervised consult the Who We Are section or co-supervised by Hjort) at our website. Cover photo: Paul A. Røstad, 1967. have been active participants. Design & layout: geriljahansen.no Hermansen, Jullum, Cunen January 2019 defended their PhDs in 2014,

2 INTRODUCTION

FocuStat: M3

M0 The M5 estimate M4 confidence curves M2 Tasks and M1 −0.0076 −0.0072 −0.0068 0.0 0.2 0.4 0.6 0.8 1.0

0 1234 0.00235 0.00240 0.00245 0.00250 Themes γ FIC To the left, a collection of confidence curves, with placebo versus drug studies. The two fatter curves in the middle relate to optimal combination of data information. To the right, a FIC plot, which evaluates and ranks different candidate models for a given purpose, related here to the slimming of minke whales over time. The more to the left, the better is the candidate model.

STATISTICS IS THE science of elling for new types of analysis, and us in the same sub-project, we have reaching decisions under uncer- potentially also for new concepts of benefitted strongly from our inter- tainty and is in many respects a information and inference. Themes national workshops and research far-ranging success story, permeat- and challenges we have addressed kitchens. We have had such theme- ing nearly all substantive sciences share the concept of the focus, the based three-day workshops during and areas of society where data operational view that the science and the spring semesters of 2015, 2016, are collected. It has used around context drive the most important 2017, with about 25 participants for hundred years to reach its present questions which again should influ- each, and smaller-scaled three-day state of high maturity and uniform ence the optimal combinations of research kitchens during the autumn usefulness. In broad strokes, the models, their analysis, and the ensu- semesters of 2014, 2015, 2016, 2017, four main areas associated with ing decisions. Three such challenges 2018, with up to a dozen partici- are as follows: pants for each. Also, in May 2018 (1) parametrics (models indexed we organised a somewhat bigger by low-dimensional parame- A: “breaking the wall” between and broader four-day summing-up ters), areas (1) and (2), partly leaning conference, with about 45 scholars (2) nonparametrics (models with on recent advances inside area taking part. high- or infinite-dimensional (3). parameters), B: extending the current scope THE PRESENT REPORT sum- (3) assessment and selection of and catalogue of approaches marises and documents some of models, and methods of relevance to our efforts and achievements. More (4) combination of different area (4), including the use of information and further details sources of information confidence distributions. may be found at the FocuStat web- C: extending and developing new site, which will be kept active and drive most of modern statistics, and methodologies for areas (1) and growing also after the formal end have, in essence, been well sorted (3) for the by now frequently of the 2014-2018 project. Impor- out, conceptually and operationally. occurring situations where the tantly, our project has involved There are important gaps to be filled number of measurements per work on two fronts, so to speak, and new paradigms and principles to individual exceeds the sample with efforts directed both towards develop, however, when faced with size. developing methodology and the statistical challenges of the 21st towards applying these for substan- century. New types of data, related THE RESEARCH AREAS alluded tive questions in real-world appli- both to new types of substantive to here are wide and the soil is rich, cation areas. Communication, the questions in a changing society and with many fruitful tasks worthy task of conveying and interpreting to evolving technologies for monitor- of detailed study, regarding both both methodology and application ing and examining more complicated methodology and applications. In stories to a wider-than-statistics phenomena than earlier, create a addition to our individual efforts, audience, has also been a crucial need for new types of statistical mod- sometimes involving two or three of dimension.

3 Focused Statistical Methods

Our FocuStat efforts have had at least two general but correlated dimensions and directions: developing and extending methodology, on one side, and applying the machinery for vital and fruitful application stories, on the other. In this section we give brief one-page summaries of certain methodological endeavours we’ve contributed to, in different focused directions.

Some of our journal publications by necessity contain mathematical theorem-proof parts, for the specialists, but here we aim at conveying the basic motivation and usefulness of our ideas. These include model building and model selection, confidence curves, change points, personalised medicine, copulae for multivariate dependence, cure models, and robustness. FOCUSED STATISTICAL METHODS

Céline Cunen: Confidence Curves for Dummies

SOME FOCUSTAT MEMBERS confidence may be read off too; and

have devoted efforts to developing we see the clear asymmetry in how 1.0 and popularising the concept of the confidence is distributed. In many confidene curve. The confidence other cases the confidence curve curve is an inferential summary. is symmetric, but here the distri-

It presents the main result of a bution of the statistic informing confidence statistical analysis, for each of the us about the size of σ has a skewed 0.0 0.2 0.4 0.6 0.8 parameters we might be particu- distribution. The cc(σ) can also be 2 4 6 8 larly interested in, and constitutes used to read off p-values for tests. σ a foundation from which the user Confidence curve for σ. We read off the median confidence point estimate can draw real-world conclusions to AT THE BASIC level the confi- at 3.226, along with e.g. a 95 percent quantitative problems. dence curve is simply a plot of confidence interval, [1.879, 7.380]. Intervals at other levels of confidence nested confidence intervals at all can be read off too. ALL QUANTITATIVE RESEARCH- levels, from 0 to 100 percent. The ERS are familiar with point plot conveys a lot of information estimates, confidence intervals, quickly and is easy to understand. THE CONFIDENCE CURVE con- and p-values. A point estimate is Some users will interpret this as a stitutes a step away from classical an intelligent guess of the correct distribution of likely values, after frequentist testing, where the typ- value of the parameter in question; seeing the data; it is the canonical ical result is simply a decision on a confidence interval gives a set of frequentist cousin to the Bayesian’s whether to reject the null hypothe- likely values; a p-value allows the posterior density (but there is no sis or not, based on some 0.05 type user to conclude a hypothesis test. prior camel to swallow). It may also significance rule. The confidence The confidence curve offers these have a positive probability mass at curve promotes transparency: things, and more. a boundary point of the parameter given the same model, data, and space, e.g. at zero for a variance assumptions, different users should HERE’S A SIMPLE example. parameter. arrive at the same confidence You observe the data points 4.09, curve, but might nonetheless draw 6.37, 6.89, 7.86, 8.28, 13.13 from CONFIDENCE CURVES ARE different real-world conclusions a normal distribution and wish particularly fruitful in the context in the end. At any rate the use of to assess the underlying spread of meta-analysis, where several confidence curves should discour- parameter, the standard deviation sources inform on the same param- age the automatic use of statistical σ. The confidence curve, say cc(σ), eter of interest, see the figure on methods and hopefully encourage sums up what can be told from the page 3. There we display a confi- critical thinking. data about that particular param- dence curve for the parameter of eter. It has candidate values of σ interest from six different studies, ALL PROCEDURES GIVING rise on the horizontal axis and level of here denoted γ, and observe that to confidence intervals can be confidence on the vertical axis. It although they inform on the same adapted in order to construct a points to the point estimate (nat- parameter there are some differ- confidence curve. Various recipes urally), here 3.226 (which we call ences, both in the point-estimates are given in Schweder and Hjort’s the median confidence estimate); and the uncertainty conveyed by Confidence, Likelihood, Probability we may read off e.g. a 95 percent each study. In addition, we also (2016) book, along with theoretical confidence interval of likely values, display the combined confidence analyses and worked-out examples. here [1.879, 7.380], cf. the red dot- curve, reflecting the full informa- ted line; intervals for other levels of tion about γ from all six studies.

5 FOCUSED STATISTICAL METHODS

Gudmund Hermansen: Change Points in Minnesota

IN JUNE 1971 the state of Minne- 1.0 sota changed their law such that Monthly Minneapolis Public Drunkenness Intakes 1.0 public drunkenness and intoxi- cation would not be considered a crime anymore. The intention with confidence sets confidence curve Number of Intakes decriminalisation was to change 200 400 600 800 Intervention 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 how local authorities would treat 1966 1968 1970 1972 1974 1976 1978 1966 1968 1970 1.0 1.5 2.0 2.5 Year and take care of individuals with year ratio of variation alcohol and drug related problems. The number of monthly public drunk- enness intakes in Minneapolis from Confidence curves for the change point January 1966 to July 1978. Can you position, left (pointing to August 1968 TO IMPLEMENT THE regulatory see that there’s another change point as being most likely), and the ratio of changes, police were given three here, around August 1968? standard deviations, right. choices for dealing with a drunken or intoxicated person in a public Cunen, Hermansen, Hjort (2018), ACTUALLY, MANY EVENTS took place: (i) transport the person to a involving also a certain statistical place and many public debates facility equipped to give emergency model for these data, we are indeed went on, related to the public use of care; (ii) take the person home able to point to August 1968 as alcohol and drugs, and the han- if he is not endangering himself, another and less visible change point. dling of such cases by the police, in others, or property; or (iii) leave Additional contextual aspects need the years leading up to the decrimi- the person where he is found. More to be brought in when attempting to nalisation and hence drastic change details are given and discussed in explain or to find proper causes for of police rules in June 1971. Our Aaronson, Dienes, Musheno (1977). such statistical change points. The change point estimate of August With this in mind, it is important data, when carefully analysed, tell us 1968, based on the data alone, also to understand questions such as (1) (a) that something changed around happens to be the very same month how did the police react to these August 1968; (b) the relative degree when public drunkenness was changes; (2) were they able to han- of that change; (c) how certain we decriminalised in Washington D.C. dle the new responsibility; (3) had can be regarding such statements; there been any noticeable changes but nothing more, so to speak. This THESE ANALYSES USE autore- in behaviour, for the police or for is conveyed via the confidence curves gressive time series models, one public drinking, before the inter- in our figure. Additional information to the left and another to the right vention date June 1971? and understanding is required if we of the unknown change point. Our wish to go on to point (d), the rea- methods point to August 1968 as a TO ADDRESS SUCH questions, sons involved and the background likely change point, and also that the consider the numbers of monthly for what we claim is such a change standard deviation nearly doubled Minneapolis public drunkenness around August 1968. Incidentally, in then. Importantly, such statements intakes, as displayed in the figure. the article mentioned we apply our come along with assessments of how One does not need to be a statisti- methods to changes in the number certain we can be, as summarised in cian to spot the drastic change point of skiing days in Nordmarka over the the confidence curves in the figure. in July 1971, before and after the past hundred years, to changes in the What is more certain than the level regulatory changes. More refined liver condition for the Lofoten skrei, of uncertainty graphed there is that methods are required when it comes again over the past hundred years, such subtle and half-visible change to looking for potential changes in and to the change of authorship in points would be very hard to find, the time period before that. Using the world’s first novel (in Catalan, quantify, and assess, without the methodology developed in our paper from the 1460ies). appropriate statistical methodology.

6 FOCUSED STATISTICAL METHODS

Kristoffer Hellton: It’s Getting Personal! Personalised Regression Models

PERSONALISED MEDICINE possible. Let the prediction for the IS becoming an essential part of new patient be our focus para­ t modern medicine. It aims to use meter, μ0 = x0 β. When the regres- an individual’s genetic informa- sion coefficients are estimated by tion to tailor drug therapies or ridge regression, the mean squared preventive care, thus taking into error (MSE) of the estimate ^ t ^ account individual heterogeneity. μ 0 = x0 β (λ), Within this framework, a patient’s the prediction error, will be a root mean square error genes, proteins, and environment function of the tuning parameter 0.75 0.80 0.85 0.90 0.95 are utilised to prevent, diagnose, λ, and for each specific vector of 1050 15 20 25 30 35 tuning parameter and treat disease. For cancer, for covariates, x0 , the MSE curve will instance, personalised medicine have a different minimum and Three different patients and their unique root-MSE curves, demonstrat- uses information regarding a a different minimand. In other ing how the minima occur at different values of λ. patient’s tumour to plan treatment words, for given x0 there is an regimes, or make a prognosis. optimal λ(x0) which minimises The Norwegian Cancer Genomics the MSE, which again depends on Consortium has been founded β and σ; this is illustrated in the FOR ALL APPLICATIONS within to establish nation-wide use of figure. If these MSE curves can be the field of personalised medicine, individual patient genetics to guide estimated separately, for each x0 , it is key to minimise the predic- cancer treatment. we can lower the expected predic- tion error for a specific individual, tion error of each single individual. rather than just the averaged error. SUPPOSE WE HAVE observed n We have achieved this via pilot Here, individual predictions can patients with a complication yi , estimates of β and σ in the MSE determine decisions with severe e.g. increase in weight, and a vector expressions, following a two-stage consequences: if a significant of ten thousand genetic variables, procedure, and we term the result- increase in a complication risk xi , and assume the outcome to ing procedure focused ridge regres- is predicted, this could trigger follow a linear regression model sion: the fridge (Hellton & Hjort, a surveillance response, or if a t yi = xi β + εi , 2018). All other current fine-tuning treatment is predicted to have a with the regression coefficients, approaches, from cross-validation high success probability, it can be β, and independent noise with a to maximum marginalised likeli- initiated despite possible adverse certain standard deviation σ. As the hood, will only select one single side effects. Hence, our goal is to high dimension requires regulari- tuning parameter value for all develop statistical machinery mini- sation, we estimate the regression future use. Our fridge procedure mising the prediction error for each coefficients via so-called ridge will adaptively modify the selected patient, rather than the average regression, which means tuning parameter to be as optimal prediction error for all patients. ^ t −1 t β (λ) = (X X + λ I) X Y, as possible for the current x0 in involving a positive tuning question, i.e. for the new patient WE HOPE THAT our methods will parameter λ. entering the doctor’s office. By util- contribute to the development of ising parametric bootstrapping, we personalised medicine and further IMAGINE NOW THAT a new can extend the proposed focused showcase the fruitful relationship patient enters the doctor’s office. ridge procedure to more general between statistics and modern We wish to predict the outcome settings, to obtain e.g. a focused medicine. variable for his or her specific logistic ridge regression. covariate vector, x0 , as best as

7 FOCUSED STATISTICAL METHODS

Nils Lid Hjort: The Focused Information Criterion

TO ANALYSE DATA, to reach clear conclusions, to form good predic-

tions, etc., statisticians need models. 109876543 Gompertz Such statistical models give precise descriptions of how data behave,

regarding both signal and noise, or nonparametric exp Weibull structure and deviations, phrased in gamma

the mathematical language of proba- log−normal

bility theory. Models range from the estimates of median differences simple (the number of times you’ll 2.5 3.0 3.5 4.0 4.5 5.0 see 1, 2, 3, 4, 5, 6 if you roll your FIC die 1000 times), to the relatively The FIC plot compares six different models, five parametric plus the nonparametric one, asso- straightforward (linear regression, ciated with six different estimates of the median age for men minus median age for women, in Roman Era Egypt. The more to the left, the better is the model, for this particular purpose, and with the perhaps ten covariates x1, the winner is the exponential model, with point estimate 5.68 for that difference. … , x10 that have been measured and taken care of in connection with a study to understand an outcome y), THERE ARE INDEED several mering out good FIC procedures to the very complex (a hierarchical different types of model selection depends on the classes of models model with three layers and a hun- methods, for sorting through worked with as well as aspects of dred nodes at each, and where some different candidate models, in dif- the focus parameter. of the mechanisms are understood ferent situations. Among the more and others not at all). well-known methods are the AIC SOMETIMES ONE MODEL might (the Akaike Information Criterion) be best for some purposes, while IN THE BEST of cases, there is and the BIC (the Bayesian Informa- another model could be better for enough basic understanding of the tion Criterion), which in different other aims; similarly, one health physics and mechanisms underlying fashions balance overall-fit with quality prediction scheme might the data that a single good enough overall-complexity. work well for men above 65, but statistical model can be put up, another underlying model could be perhaps followed up with a good- THE FIC (THE Focused Infor- better for women between 50 and ness-of-fit test verification. In most mation Criterion), developed by 60, etc. The FIC machinery assists applications, however, particularly G. Claeskens and N.L. Hjort and in such modelling choices. if the data are complex, there will be later extended in several new many competing candidate models directions and for new purposes THE FIC PLOT here serves as a

(which of the ten covariates xj ought by the FocuStat group, works in a simple illustration, using data on to be included, and with which extra different fashion. It takes specifi- life-lengths from Roman Era Egypt, interactions?; which of the nodes cally on board the notion that some the first century B.C., for 82 men should be part of a final hierarchical questions are more important than and 59 women. We fitted six differ- structure?). To handle such issues others. It starts by (1) formalising ent models to the data, and can for statisticians need an arsenal of what is most important with the each of these estimate quantities model selection methods, followed final analysis, in the form of a con- of interest, along with measures perhaps with tools for forming text driven focus parameter, and of uncertainty. Different focused data-driven weighted averages of then (2) assessing and ranking the questions lead to different FIC estimates or predictions coming out precision of all candidate estimates plots and potentially to different of the best-looking models. for this focus parameter. Ham- rankings of the candidate models.

8 FOCUSED STATISTICAL METHODS

Vinnie Ko: Multivariate Modelling with Copulae

PROBABILITY DISTRIBUTIONS ARE without a doubt at the heart of statistics. A probability distri- bution is a mathematical function that provides the probabilities of 15 Copula density occurrences of different events, or 10 outcomes (e.g. the probability of 5 getting 6 from throwing a dice). 1.0 When the probability distribution 0 0.8 0.0 0.6 is only concerned with a single 0.2 0.4 0.4 variable, it’s called univariate (e.g. Lørenskog 0.6 Nannestad the probability of precipitation at a 0.2 0.8 certain level a in city A). When the 1.0 0.0 probability distribution concerns Copula density plot for daily precipitation data, in Lørenskog and Nannestad. multiple events of interest, or mul- tiple variables, it’s called multivari- ate (e.g. the probability that there is that its name “copula” comes from WITH THIS MODEL, one can for precipitation at a certain level a in the same Latin word, meaning instance answer questions like city A and at another level b in city “link”, or “tie”, explains its nature “Given that there is precipitation B). Most of the well-known multi- well. Once univariate distributions above 10 mm in Nannestad, what is variate distributions require that and a copula are chosen, and the the probability that Lørenskog also each of its univariate components parameters are estimated, one has has precipitation above 10 mm?”, has the same type of distribution. a model that can fully describe and “What is the chance that both This is often an unrealistic require- probabilistic relationships between Lørenskog and Nannestad will have ment in practice. For example, the events of interest. This works for more than 15 mm rain three days in probability distribution for precip- any number of variables, e.g. as a a row?”. itation in Oslo can be very different modelling tool for analysing and from that of Bergen, while they predicting the precipitation level in AS PART OF my FocuStat work, I nevertheless have a joint multivari- five Norwegian cities. have worked with different meth- ate probability distribution. ods for estimating the parameters AS AN EXAMPLE, we use a subset in such models, and also developed A COPULA IS a general statistical of the daily precipitation data from model selection criteria that help tool for modelling multivariate the Norwegian Meteorological selecting the most appropriate probability distributions. It has as Institute. We look at daily precip- copula for specific situations (see a strong advantage that it allows itation, measured in Lørenskog the Ko and Hjort papers in the its univariate components to have and Nannestad, between 1990 and References section). different parametric forms. The 2006, only for the days when there main idea behind it is to “cancel was precipitation in both cities. The out” each univariate distribution figure shows the so-called copula by transforming them with their density plot from the fitted copula own inverse. These transformed model. It basically shows that there versions are then connected with is a positive dependence relation- a copula to form a multivariate ship in the precipitation between probability distribution. The fact the two cities.

9 FOCUSED STATISTICAL METHODS

Emil Stoltenberg: Relapse after Rehab

THE FIGURE DISPLAYS the time

in days for 575 drug rehab patients 2.0 from entering rehab to an eventual relapse. The black lines belong to baseline 0.0 0.5 1.0 1.5 those that suffered from a relapse, 0 117 234 352 469 586 while the grey lines are those cumulative coefficient Days

patients for which no relapse was 0.0 observed. Standard survival analy-

sis methods – which due to the time treatment −0.4 −0.2 component and the censoring are 0 200 400 600 800 1000 1200 0 117 234 352 469 586 cumulative coefficient often employed to analyse data sets Time to relapse in days Days

such as this one – assume that all Normalised times to relapse of 575 Cumulative regression coefficients. the patients in the sample will even- drug addicts in rehab. Black lines are Baseline and whether the patient was observed relapses, the grey lines are scheduled to stay in rehab for three tually relapse. In many applications censored observation times. months (coded 0) or six months (coded 1). this is an untenable assumption (luckily). In our example, rehab works for some patients, due to THE MOST COMMON cure model length of stay in rehab, three their genetic composition or the takes the population survival months (coded 0) or six months

like, while others are doomed to a function as Spop(t,x,z)=1−π(x)+π(x) (coded 1). The probability π(x) was strenuous life of rehab-relapse-re- S(t,z), where π(x) is the probability taken to be the logistic function hab-relapse ad infinitum. This of being susceptible to the event with a covariate giving the so-called raises several statistical questions. of interest, often modelled as the Beck score for depression when First, we would like to estimate the logistic function, with x a vector entering rehab. The estimates are size of the two groups – the share of covariates, and S(t,z) is the rather bleak here; the most and of drug users for which rehab might survival function of the susceptible the least depressed have estimated work, and if data are available, what individuals, that also may depend probabilities of 0.99 and 0.88 of factors might contribute to an indi- on covariates, typically via a Cox relapsing, respectively. On a mod- vidual belonging to this group. Sec- regression model. An alternative estly brighter note, the estimated ond, survival estimates such as that to the Cox model is Aalen’s linear cumulative regression coefficient of the risk of relapsing must take hazard rate regression model. It on the treatment indicator indi- into account that some drug users postulates that the i-th individual cates that a longer stay in rehab has are cured, i.e. will never relapse; if has a hazard rate function being a a protective effect, compared to a not, they are likely to underestimate linear combination of covariates shorter stay.

this risk. and regressor functions, say αj(t). THERE IS GOOD scope for A CLASS OF survival analysis mod- IN SOME ONGOING work I extend extending and finessing methods els suited for such situations takes Aalen’s model for the hazard rate pointed to here, using elements of the population survival function, to the standard cure model above. the FocuStat machinery. One may affecting both groups, as improper. This involves certain cumulative construct parametric models for

That is, Spop(t)=P{alive at time t} regression functions. The figure some of the αj(t) functions; it would tends to a number bigger than zero displays the baseline cumulative be useful to develop FIC schemes as time goes on. In our example, regression coefficient and the for selecting models best suited to this is the share of cured patients cumulative regression coefficient the cure fraction; etc. for whom rehab actually works. on the indicator of the planned

10 FOCUSED STATISTICAL METHODS

Sam-Erik Walker: Robust Estimation

ALL STATISTICAL METHODS use data. Most often these are all good and the statistical method yields satisfactory results. But sometimes the data might contain atypical or even incorrect values or entries, Median Median which should then ideally be Models Models Pareto Pareto Inv. Pareto Inv. Pareto spotted and removed before further Inv. Burr Inv. Burr 4000 6000 8000 10000 12000 3500 4000 4500 5000 5500 analysis. Running an error-re- 2000 3000 4000 5000 6000 1000 1500 2000 2500 3000 3500 4000 silient approach with automatic FIC FIC screening and removal of atypical data without a proper analysis is FIC plots for the median for pre-1950 data (left) and post-1950 data (right) for the Pareto (green), inverse Pareto (orange), and inverse Burr (red) models, using the BHHJ estima- not recommended, however. Even tor with a range of the tuning parameter a. The triangle identifies the maximum likelihood. if a reasonably thorough check The nonparametric estimates are shown as the blue points. is carried out, one can easily be misled by deciding to down-play or ignore such data. A good example figure provides an illustration of based on pre-1950 and post-1950 of this was in the slow discovery of this, relating to the estimation of war data. In each plot, a FIC curve is the ozone hole over Antarctica in the median number of battle deaths given for each of the three paramet- the mid-1980s (read Pukelsheim, before and after 1950 based on the ric models, describing a relationship 1990); for several years one saw the Correlates of War database (Cunen, between the accuracy of the median atypical but genuine data as errors. Hjort, Nygård, 2019, Walker, Hjort, estimates, via the root FIC score on 2019), which contains the number the x-axis, and the resulting esti- THERE ARE ALSO various robust of battlefield deaths for the 95 mate of the median on the y-axis. estimation and analysis meth- biggest interstate wars, from 1823 Each FIC curve is plotted for a rele- ods which do well even without to the present. vant range of a. The best combina- such initial data inspection work. tion of model and robust estimator These are methods specifically FOR THIS ILLUSTRATION we is obtained by the point on a FIC constructed to produce reasonably take the median as the focus curve lying most to the left. For good results even in the presence of parameter, and FIC plots are the pre- and post-1950 data this is bad data, or, more generally, when shown involving four models obtained for the Pareto model, with the model used is not fully correct. (Pareto, inverse Pareto, inverse certain best balance parameters, Burr, and the nonparametric), and with best estimated median val- ROBUST METHODS CAN also be where the parameters of these ues of 7091 and 4272 battle deaths, combined with so-called focused sta- models are estimated using a cer- respectively. We learn that robust tistical approaches, one of the central tain BHHJ estimator (Basu, Harris, use of a parametric model leads to themes of the FocuStat project, Hjort, Jones, 1998), with its robust- the best median estimates. e.g. via appropriate finessing of the ness vs. efficiency trade-off param- Focused Information Criterion, or eter a≥0. Here a=0 corresponds to THE FIC METHODS of Walker and FIC. It can be used to select among the non-robust maximum likeli- Hjort (2019) used here represent models based on a focus parameter, hood estimator, and increasing a an extension of those developed a certain parameter or function of implies more robustness. in Jullum and Hjort (2017, 2018), parameters considered most import- from maximum likelihood to ant in a given statistical setting. The THE LEFT AND right plots are classes of robust estimators.

11 WORKSHOPS AND THE VVV CONFERENCE

May 2016: FICology Statistical Stories

Our project has had significant methodological components, resulting in general analysis machineries for different purposes of statistical inference. We’ve had Mention the good a fair share ofMention application the good stories, though. Mention the good things told to us by things told to us by things told to us by our participants.These have involvedour participants. wars, whales, angels, armed our participants. conflicts, medical statistics, speedskating, Game of Navn/tittel/whathaveyou Navn/tittel/whathaveyou Navn/tittel/whathaveyou Thrones, the world’s first novel, the longest time series ever from fisheries science (with the Hjort index), the cooling of newborns after oxygen deprived births, air pollution, vocabulary sizes for toddlers, Ngrams for IBEARUM QUI SUS esciisttext ectorer analysis, why Americans chose to elect Mr Trump, uptaerem endentus eictaepFrench eribus grammar, macro-economy, boy-girl analysis, eum eniandae sinctibus andquia many more. Here we offer brief one-page statistical volupta epuditiore voluptat perorio stories from our desks. omnit velitio te nus sum sent assiminci andes magname necus.

LUM, OD QUE cuptis nonseq- uis est, qui res mi, sollecero ent es exerum endelestrum iminver cillaut hicipsu ntempore sum ut inveliquid quid ut atiscias quae. Ut exereperit, sam vellam nossent ere, in es sequam, sapernam quidit vitis dolendiam fuga. Namet min renime cus ullento beatio quid modit, omnitibus,

12 STATISTICAL STORIES

Céline Cunen: Are the Antarctic fat weight 0.0 0.2 0.4 0.6 0.8 1.0 Minke −0.015 −0.010 −0.005 0.000 β Whales 1.0 1.5 2.0 2.5 51015 Confidence curve for the linear yearly decline. We read off a natural point year Becoming estimate at -0.008 and a 95 percent Fat weight (in tonnes), for 742 whales, confidence interval, [-0.014, -0.002], Thinner? per year, over 18 years. with the scale being in tonnes.

THIS INNOCUOUS AND straight- in krill production, the abundance several explanatory variables (both forward-sounding question has of other whale species, or climate fixed and random, in particular been at the centre of a decade-long change. These potential explana- the year-to-year variation), along debate in the Scientific Commit- tions are fascinating, but first, one with interactions. The focus of our tee of the International Whaling has to establish whether there is investigations is the overall yearly Commission (IWC). Set against a any change in fat-reserves at all, decline in fat weight, related to backdrop of political controversies, and estimate the magnitude of that the parameters of our candidate the debate itself concerned biologi- change. Through our contact and models. We have extended the FIC cal, and primarily, statistical issues. collaborator Lars Walløe, we had machinery to this class of models, These issues led to the involvement access to measurements of body thus helping us finding the best of some members of the FocuStat condition on hundreds of Antarc- model for this particular question. team. tic Minke whales in a period of 18 years (from 1987/88 to 2004/05). OUR RESULTS WERE presented THE POLITICAL ASPECTS of the One of the body condition mea- in the form of a confidence curve debate arise from controversies surements is the fat weight of each for the yearly decline in fat weight. surrounding Japanese Scientific whale, which we display in the The curve indicates an effect which whaling. The data involved here figure above. is negative and significant. The originate from such whaling, the magnitude of the point-estimates Japanese Whale Research Pro- THE FIGURE DOES not give the corresponds to the average whale gramme under Special Permit in impression of a clear decline in experiencing a loss of about 80 the Antarctic (JARPA). The con- fat weight. But wait! The differ- kg of fat over ten years. A typical troversies and discussions in the ent whales in the dataset are not whale has between one and two IWC are certainly interesting, but directly comparable to each other; tonnes of fat, but bear in mind outside the scope of this small note. some are male, other female, they that the fat-reserves are vital for They have been treated in some are caught in different periods of the whales’ reproductive success length in our blog post Whales, the year, and at different latitudes. and survival in the winter (when Politics, and Statisticans. Here we In order to correctly address the the whales typically do not feed). will focus on the statistical analyses question of decline in fat weight, We concluded that the Antarctic and the results. a statistical model needs to be Minke whales had indeed suffered built. The model is constructed a loss of fat-reserves from 1987/88 BIOLOGISTS ARE INTERESTED by considering the various factors to 2004/05, and after several days in the fat-reserves of Antarctic which contribute to the observed of heated debate, our conclusions Minke whales because large-scale fat weight of each whale. Here were essentially accepted by the changes in such quantities could we used several candidate mod- Scientific Committee of the IWC. herald deeper transformations in els from the class of linear mixed the ecosystem, related to changes effects models, involving a list of

13 STATISTICAL STORIES

Gudmund Hermansen: Armed Conflict Colombia Non−State One−Sided 0 500 1000 The Dramas 1989 1992 1995 1998 2001 2004 2007 2010 2013 2016 Simulated Colombia Battle deaths of Escalation 0 500 1000 0 1000 2000 3000 1970 1980 1990 2000 2010

Year 0 500 1000

The number of battle deaths in the The sad Colombian truth (top panel), Colombia conflict. Solid line: armed followed by two simulations from conflict, with the state (Colombia) the estimated statistical violence one of the actors; dashed line: two process model (two lover panels). In non-governmental actors; dotted detail, the plots display the quarterly line: one-sided violence (by the number of battle deaths in Colombia, state, rebel, or terrorist group). from 1989 to 2016.

WARS ARE BEING fought in Syria, study of political and social order, a higher level. A fruitful approach Iraq, Afghanistan, Somalia, with and for preventing the human suf- is in terms of an underlying conflict these and other ongoing wars and fering that escalation inflicts. intensity, which might evolve violent conflicts producing death dynamically over time, and depend- tolls not seen since the Cold War. WE MIGHT ALL have an intuitive ing on relevant covariates (these Existing instruments, from diplo- notion of what escalation or conflict include economy, natural resources, macy to power politics, have largely escalation means, but this key con- type of government, number of proven inadequate for stopping or cept is elusive and difficult to define fighting neighbours, etc.). The reducing the violence in such con- and model. We take observed escala- ­number of fatalities are then flicts. Furthermore, policy makers tion to mean a substantial change in Poisson- or mixed Poisson-type as well as the research community the lethality, geographic scope, num- outcomes, stemming from the lack proper understanding and ber of participating actors, targets of underlying conflict intensity func- good answers to several important violence, repertoires or technologies tion. Our escalation models take questions. Why do some civil wars of violence utilised, or aims of the on board the possibilities for rapid remain low-intensity conflicts, actors involved in conflict; cf. what is changes, and relative constancies whereas others escalate, to claim the graphed in the figure. More refined for periods of time. That our mod- lives of hundreds of thousands? And definitions are required, however, for els work is reflected in simulations, why do some non-violent protest a more systematic understanding; with spikes and plateaus, see the movements escalate into bloody policy makers rely on insights from figure. Such may also be used to civil wars, whereas others remain high-quality research studies. predict future levels of conflict peaceful? There are no clear answers under different scenarios. to the key questions of how, why, THE CONCEPTUAL AND statisti- and when do low-intensity conflicts cal definition of escalation should AN IMPORTANT ASPECT of our experience escalation in violence, as depend on which questions we aim modelling work is that a certain in the case of Colombia between the to answer. In ongoing work with increase in violence level should not government and FARC, which saw Håvard Nygård (PRIO), we have too automatically be interpreted as a ten-fold increase in battle deaths pursued several paths, for building a genuine escalation (in contrast from 1993 to 2005, after thirty years meaningful statistical models for to what is seen in other quantita- of relatively low-intensity conflict. the violence processes. These should tive analyses of these conflicts). capture the essentials, in terms of Among the aims, for ongoing and CONFLICT RESEARCHERS HAVE inputs, reflecting different scenar- planned work, also for the Stability only a rudimentary understanding ios or decisions, and simulations and Change project with Hjort and of the determinants and dynamics from an estimated model should Nygård, is to move beyond the cur- of conflict escalation. A deeper look like “the real thing”.One might rent understanding of escalation, in understanding of conflict escalation build models for the probabilities of order to obtain the level of knowl- is critical both for advancing the passing from one level of conflict to edge required to bring about change.

14 STATISTICAL STORIES

Kristoffer Hellton: I Have a Dream! The Drug Synergy DREAM When combining two drugs in a treatment regime, the naive result is that they work inde- pendently and give an additive, doubling of the effect. If the effect more than doubles, the Challenge drugs have created a synergy. In the worse case, the drugs are antagonistic and result in a less than additive effect.

IN THE AUTUMN of 2015, I tions copy-number alterations, and used to predict the monotherapeutic joined a Big Insight team entering methylation data, accessible to all effect of each drug, based on the the AstraZeneca-Sanger DREAM participants. The main challenge drugs chemical and molecular char- Challenge, a community effort to was to develop statistical method- acteristics. Then a multiresponse develop new approaches of predict- ologies which would control the ridge regression approach was used ing beneficial drug combinations. large combinatorial space of poten- to predict synergy based on all the Targeted therapies and personalised tial combinations and at the same different available genetic data. treatments are the most promising borrow information across both Unknown synergy values in the test assets for treating cancer. However, drugs, cell lines and the range of set were filled in following the iter- in many patients, a tumour can be genetic data. The available data set ative Healy-Westmacott procedure. innate or have acquired a resis- was partitioned such that 2/3 could Both models were further fine-tuned tance to given therapies, which can be used for training the models and by cross-validation. In the end, all render treatments ineffective. To the remaining 1/3 needed to be models, also the separate ­Cambridge increase therapeutic options, cancer predicted as correctly as possible in and Oslo contributions, were researchers have been actively the challenge finale. averaged to produce an ensemble investigating new drug combina- prediction model with weights based tions. When combining different THE OSLO TEAM, organised by on cross-validated prediction errors. treatments in new ways, the most Arnoldo Frigessi and Manuela desirable property is drug synergy Zucknick at the Department of WORKING IN A fast paced team – an amplified response over and Biostatistics, formed a joined team, with a high energy spirit was beyond an additive effect. Drug syn- minor insight, together with the both immensely fun and scientif- ergism has the potential to increase MRC Biostatistics Unit at Cam- ically rewarding. The Cambridge anti-tumour potency without bridge, led by Lorenz Wernish and participants even visited Oslo for increasing toxicity, but it is very com- Sylvia Richardson. To best exploit a a few highly enjoyable days of plex to study due to the enormous final model ensemble step, we fol- intense interdisciplinary teamwork. number of potential combinations. lowed different but complementary Admittedly, we lost to various other modelling strategies. The Cambridge teams, but gained vital insights FOR THE DREAM challenge, team utilised kernel radial basis from our modelling efforts. In AstraZeneca released around 11 regression with a ridge penalty, retrospect, we have concluded that 000 experimentally tested drug based on constructing similarity our proposed models were too rigid combinations for 118 drugs (with matrices between all observations and at the same time did not incor- over 7000 combinations) in 85 with all data pooled together. The porate enough genetic information. different cancer cell lines. More- Oslo team had separate approaches The community effort of advancing over, the Sanger Institute made for the drug characteristics and the drug synergy prediction is currently measurements of genomic data, genetic data. First, a gradient boost- being published as a consortium including gene expression, muta- ing model with non-linear trees was paper in the journal Nature.

15 STATISTICAL STORIES

Nils Lid Hjort:

War, 6 2141 0 1 2 3 01 81 log battle deaths Peace, and −1 quantile change score Statisticians 1850 1900 1950 2000 1850 1900 1950 2000 history history

log-battle-deaths for the 95 major inter- Two checking-for-constancy plots, for state wars, from 1823 to the present. the median and for the 0.75 quantiles, The horizontal red line indicates the both relating to measuring the relative level above which there is a so-called degree of change, to the left and to power-law statistical structure. the right of each change-point candi- date position along the time axis.

EACH WAR IS dramatic and warring nations, registred before depends on how questions are horrible and warrants a hundred the outbreak of war. posed and finessed; it is too much history books and a thousand songs to hope for a very simple “the world of sorrow. But it also provides one SEVERAL MAJOR QUESTIONS changed for the better in 1950, more data point, along with rele- can be posed and analysed just period”. Indeed, when we shift our

vant covariates, to the collection of with the (ti, zi) data, cf. the first attention to the bigger wars, sepa-

Tolstoyan war-and-peace data- figure, which shows (xi , log zi) rately, those at so-called power-law bases, and statisticians may study over time. In some of our analyses behaviour level, our methods point the evolution of alleged decreasing we demonstrate that the world of to 1967, the time of Vietnam and violence levels over time. One of wars has not remained constant Flower Power. the bigger questions is both decep- over time; there is a statistically tively simple and quite controver- significant change of levels, for the SO WE HAVE contributed to the sial: is the world becoming more better. The second figure is a case statistical verification of The long peaceful over time, or not? in point, where we show, via spe- peace. But what are its contributing cialised models and methods that factors, and what political aspects IN HIS FAMOUS book The Better we have developed, that any change might be at stake, in order to avoid Angels of Our Nature, Steven Pinker in the median of the distribution is future escalation? This is where the has a long chapter on The long too smal to register properly, but covariates come in, the sequence of

peace, and argues, along several that there is a significant change if xi (along with delicate discussions paths and dimensions, that the we focus on e.g. the upper quartiles of precisely what to put into these world is become steadily more (the 0.75 quantile). Also, the esti- scores, how to measure democracy peaceful. There are fierce voices mated regime-shift, or historical level, etc.). In addition to securing against such a verdict, however. In change-point, is at 1950, the time high-quality data of relevance, one the Cunen, Hjort, Nygård article of the Korean War. The red curve needs more advanced statistical Statistical Sightings of Better Angels relates to comparing medians, and models and methods. Our paper (2019), we report on several rele- the flucuations around zero are not delves into such questions too. We vant statistical models and analyses, significant. The black curve relates find that having a high democracy pertaining to data which we in this to comparing 0.75 quantiles, where score is associated with a clear

simplified discussion denote (ti, xi , the plot signals (a) that there is a decrease in battlefield fatalities.

zi), for the 95 major interstate wars significant change, and (b) that The broader hope is to understand from 1823 to the present. Here ti is the likely year for that change is in at a deeper level, both statistically the onset time, zi is the battle death 1950, the Korean war. and in terms of political processes count, and xi contains relevant and actions, what leads to stabil- covariate information; we have e.g. AS THIS SHORT glimpse into the ity, or to escalation, or to war. Our utilised a certain democracy score, modelling and analyses already models may also be used to predict to form such a covariate for the indicates, what we are able to find aspects of future conflicts.

16 STATISTICAL STORIES

Vinnie Ko: 234 focustatske meldinger Vinnie Ko's 2016 book delivers a multitude of both humorous and serious reports on attempting to understand the Dutch, the peo- ple and the language, seen through Korean eyes. With a modicum of trepidation we look forward to his book on how to understand Norway (if possible), her language and her statisticians, seen through Korean and Dutch eyes.

I 2016 FLYTTET jeg fra Nederland window for our weekly meetings – DETTE VAR IKKE begrenset til til Norge for å begynne med min so find your calendars.» faglige punkter, men handlet om doktorgrad på UiO. Jeg var den alt som hadde skjedd med oss heldige, med et stort privilegium, I DET ØYEBLIKKET var jeg i FocuStat: «Vinnie har kjøpt ikke bare fordi jeg fikk lov til å nettopp ferdig med mitt seksukers smågodt til prisen kr 30 pr kg.» begynne et nytt eventyr i et nytt norskkurs, hvor vi hadde øvet med (nr. 363, 24.10.2017) Og noen land, men jeg hadde også total fri- ting som «Jeg liker meg her i byen» ganger handlet det om nødvendig het til å velge hvem som helst som og «Jeg vil gjerne ha en pølse i tvangsarbeid: «Folkens, ta heisen veileder for min PhD, siden min lompe». Jeg tenkte: sånn bruker opp til 7. etasje, til kl 18:29:59. finansiering kom fra Matematisk ekte nordmenn sitt språk! Og skrev Så tar vi noen tvangsfotografier institutt selv. ned ‘hilseinterjeksjon’ og ‘kolli­ der, hvorpå vi kanskje også unner sjons­matrise’ i ordlisten min. oss å spasere til ‘Focus på torvet’, ETTER AT JEG hadde snakket fire minutter unna, for å tvangs­ med «alle de voksne som muligvis FRA MØTE NR. 2 bestemte Nils seg smile enda litt mer.» (nr. 332, kunne være veileder» innenfor for å kjøre alt på norsk igjen, akkurat 14.06.2017) Eller om konsekvenser: statistikk-seksjonen, valgte jeg som før jeg kom i gruppen. I de «Hvis vi ikke får godkjent vår Nils Lid Hjort, hovedsakelig fordi første månedene kunne jeg ikke følge statusrapport, legges prosjektet han har kule og brede interes- alt hva de seks norske sa. Da var øyeblikkelig ned, og dere settes på ser, og forståelse for hva jeg gjør det de focustatske meldingene som gaten (men ikke jeg, da).» (nr. 254, utenfor statistikk. Etter vårt første reddet meg. Før møtene sendte Nils 25.09.2016) møte, foreslo han at jeg ble med oss melding om kommende møte og på hans FocuStat prosjekt. Et par saker. Og etter møtene skrev han en I BEGYNNELSEN MÅTTE jeg søke dager senere fikk jeg min første e-post om alt vi hadde gjennomgått, ord og uttrykk i ordnett.no hver focustatske melding, adressert til og med tanker om det neste. halvsetning for å dechiffrere ordene alle syv deltakere av gruppen: hans. Men takket være de foreløpig Dato: 24.08.2016 «RUNDE RUNDT BORDET» hadde 234 focustatske meldingene, ble jeg Emne: focustatsk melding nr. vi stadig: «Kristoffer forteller fra herre over norsk, med svært høyt 246: ny møtetid. «Hei (der denne sin eksotiske reise; Sam-Erik om de stigningstall i læringskurven. kollektive hilseinterjeksjon nå siste syv metre frem mot comple- også omfatter Vinnie), [...] Vinnie, tion for Two; Gudmund om diverse SNART NÅR FOCUSTAT sin ende, kan du maile til meg og Celine, og Data Science kick-off; Nils om og vi skal levere inn en sluttrap- med opplysning om *which hours siste fire gode ideer innenfor chan- port. For denne gangen er jeg ikke during the week you're busy with gepointeri, og om Kongen igår; redd for å settes på gaten. Hvis lectures or other matters*? After Celine har knapt med tid, med to vi skriver ut alle 478 focustatske the required kollisjonsmatrise foredrag og tre timers Bayes hver ­meldinger og stifter dem sammen, exercise we'll then find a new time fredag …» (nr. 450, 05.09.2018) har vi allerede en fin sluttrapport!

17 STATISTICAL STORIES

Emil Stoltenberg: Graduation Graduation + 1 Graduation + 2 Boys, Girls,

1 2 3 4 5 6 21 3 4 5 6 21 3 4 5 6 and Gauss Average grade Average grade Average grade

Graduation + 3 Graduation + 4 Graduation + 5

1 2 3 4 5 6 21 3 4 5 6 21 3 4 5 6 Average grade Average grade Average grade

Boys catching up to girls after school age. Hypothesised normal densities for girls (solid curve, sd=0.50) and boys (dotted curve, sd=0.50). Difference in mean going from 0.048 to zero.

IN THINKING FAST and Slow, education in Norway, 500 girls and out. Then 31% of the boys and 17% Daniel Kahneman writes about sta- 500 boys, all dreaming of spending of the girls drop out (not too far tistical intuition. A person, he writes, at least five years at the University from the actual figures). Clearly, one can possess statistical intuition albeit of Oslo. The UiO only accepts 200 reason for “dropout-boys doing bet- being devoid of mathematical skill. students, and only cares about the ter than dropout-girls” might just I agree. The following hypothesis, average grades of the applicants. be that they outnumber the girls, which is not mine, but belongs to my Assume that the perfomance of the so just based on the numbers we'll mother, is a good example of how girls and the boys can be repre- hear more stories about the former statistical intuition can guide think- sented by independent draws from dropout-boy now being his own ing about a social phenomenon. normal distributions with the same employer (en gründer!), and we are spread, but differing means. Set the led to think that dropping out is less THE SHARE OF men with a fragile standard deviation to 0.50 and the serious for boys than for girls. connection to the labour force, who means to 3.37 for the girls and 3.13 do not find a partner, nor have for the boys; see upper left in the IT MIGHT ALSO be the case that any children constitute a growing figure. So how many girls and how dropping out of school has less minority in many western countries. many boys will commence studies at serious long-term consequence for As with all social phenomena, this the UiO after the summer? Among boys than for girls. Consider the might be due to a plethora of reasons. the 200 students accepted we can normal densities in the figure. The One reason, nevertheless, seems to expect there to be 140 girls. This five subsequent plots illustrate how be that girls perform better in school number, 70%, is close to the actual boys might catch up with the girls than their male peers. And the reason at the prestigious medicine and law as time proceeds – think of each why the girls outperform the boys programmes in Norway. plot being one year extra removed might simply be that girls mature from graduation – while the girls, earlier than boys. Instead of searching THERE IS A perceived social in our toy example, are fully mature for complicated societal explanations, phenomenon for which our normal at graduation. Under this scenario, my mother's statistical intuition was distribution toy example can about half of the boys who dropped that the normal distribution should provide insights: among students out of school would not have done have something to tell us. who do not finish upper secondary so five years later. The drop- education, the dropout-boys seem out-girls, on the other hand, would IN NORWAY IN 2017 the average to fare better later on in life than the still have dropped out five years grade (grades range from 1 to 6, with dropout-girls. Again, we have two later. In this way, the delayed mat- 6 being the best) in Videregående normal distributions with means uration of boys might explain why skole was 3.37 for girls and 3.13 for 3.37 and 3.13 for girls and boys, and dropout-boys appear, or indeed boys. Let's simplify. Suppose that the same standard deviation 0.50. are, more prone to succeeding in in 2017 there were 1000 aspiring Assume now that the students with life than the girls who shared their students finishing upper secondary an average grade below 2.88 drop fate as late teens.

18 STATISTICAL STORIES

Sam-Erik Walker:

beta −2.50 −2.25 −2.00 60 −1.75 −1.50 −1.25 −1.00 −0.75 −0.50 −0.25 55 0.00 0.25 Ozone 0.50

50 confidence Decline 45

40 0.0 0.2 0.4 0.6 0.8 1.0

35 −0.45 −0.40 −0.35 −0.30 −0.25 −0.20 over −10 −5 0510 15 20 beta Europe Estimated trend coefficients of summer Confidence curves for the overall trend season ozone from GAM modelling, for coefficient β0 of the summer season the 302 rural stations, for the period 2000 ozone in Europe, at the rural (black, to 2010. Negative and positive trends left) and suburban stations (red, right), are shown using blue and red colours, for the period 2000 to 2010 (left plot). respectively. The unit is ppb/year. The unit is again in ppb/year.

HAS THERE BEEN an overall from 2000; finally, ε(t) represents AS SEEN HERE, the confidence decline in surface ozone in Europe, the error term, taken normally curve to the left points to an overall over the past decades? The answer distributed with mean zero. The average decline in summer season turns out to be yes, as we report on linear trend may be interpreted as a ozone in rural Europe of -0.391 ppb here. The data consist of maxi- meteorology adjusted trend at each per year, or -3.91 ppb per decade, mum 8-hour running daily average station, mainly influenced by trends with [-4.33, -3.50] as a 95 percent concentrations of ozone, for the in various precursors to ozone from confidence interval for the latter. summer seasons (1 May to 31 Aug), emissions of nitrogen oxides (NOx) For the suburban stations, the at 302 rural and 210 suburban and volatile organic compounds overall estimate is -0.281 ppb per stations, across Europe. (VOCs) from traffic and industry. year, or -2.81 per decade, with 95 percent interval [-3.40,-2.22]. This AT EACH STATION, a specially OUR FOCUS HERE is on the represents a small but significant designed Generalised Additive trend over time, specifically the β decline in the ozone levels corre- Model has been fitted to the ozone parameter. This trend parameter sponding to around 0.8 percent data, using corresponding mete- varies across Europe. Analysing the per year, or 8 percent per decade, ^ orological data at each station as model above yields an estimate β i relative to the typical mean level of smooth explanatory variables. The for station i, along with a standard summer ozone at the rural stations, meteorological data consist of mod- deviation σi. As seen from the fig- for the period 2000 to 2010. The elled daily values of air tempera- ure, a negative (downward) trend is decline for the suburban stations is ture, relative humidity, wind speed, found at most of the rural stations, about 6 percent per decade. Since solar short-wave radiation and but not all. the individual GAM based trends at atmospheric boundary layer height. the stations were meteorologically In addition, a day of season variable THE GRANDER QUESTION is adjusted, this decline is probably (relative to 1 May each year) and a whether the overall European ozone mostly due to a reduction in emis- linear long-term trend describing trend has been negative, for the sions of precursors to ozone such the potential change in ozone over 2000 to 2010 decade. We answer as NOx and VOCs from traffic and time were introduced as additional this by modelling the individual industry. This is good news, since explanatory variables. Thus, at each βi as coming from an appropriate surface ozone represents a poten- station, the model is background distribution, a normal tial hazard for both health and 2 O3(t)=α+βt+∑jsj(xj(t))+ε(t), N(β0, τ ). Using modern meta-anal- agricultural crops. relating the daily ozone values ysis, via confidence distributions, as

O3(t) to smooth functions sj(xj(t)) developed in Schweder and Hjort ELSEWHERE WE WILL also of the meteorological and day of (2016, Ch. 13), we can estimate both report on the degree of variation, season variables xj(t) and to the the overall trend parameter β0 and for this negative ozone trend crucial linear trend in ozone α+βt, the spread parameter τ, along with parameter, across Europe, also for where t represents time in years, confidence curves. the suburban data.

19 Workshops Programmes: and the VVV Conference

We have organised theme-based three-day international workshops in 2015, 2016, 2017, with about 25 participants for each, and a somewhat bigger four-day international Inference With Confidence summing-up conference in 2018, with about May 2015 45 scholars taking part.

Planning and organising these have been tasking, challenging, time-consuming, but also fun, fruitful, and inspiring. They have involved and led to collaborative work and fostered new ideas. The three three-day workshops were held in Teknologihuset, close to Bislett, inspiring also Bislett-related cover FICology photos for our programme booklets. May 2016

Building Bridges at Bislett May 2017

Vårens Vakreste Variabler May 2018 Gruppenbild mit Dame, from one of our workshops at Teknologihuset. WORKSHOPS AND THE VVV CONFERENCE

May 2015: Inference With Confidence Confidence Distributions and Related Themes

›› The Holy Grail ›› Posteriors without priors 3 23 16 7 ›› Optimal inference, days participants talks countries with confidence

IN MAY 2015 we did Inference With Confidence, on Confidence Distribu- tions and Related Themes.

SPEAKERS WERE C. CUNEN, A. Frigessi, S. Grønneberg, J. Hannig, K.H. Hellton, G.H. Hermansen, N.L. Hjort, B.H. Lindqvist, R. Liu, T. Schweder, D. Sun, G. Taraldsen, P. Veronese, S.-E. Walker, M.-g. Xie. Themes were partly of a theoreti- cal nature, developing confidence distributions for new purposes, but also included real-world application stories. The Excursion With Confi- dence was at Ekebergparken, with a forestful of sculptures.

OUR WORKSHOP EFFORTS led also to a Special Issue of the Journal of Statistical Planning and Inference, with N.L. Hjort and T. Schweder as guest editors for the occasion (assisted crucially also by C. Cunen’s efforts). This actually involved almost two years of editorial pro- cessing, with papers to be finished and then contributed, reviewed, sent back for improvements, etc., before the final versions appeared in the journal (Vol. 195, May 2018, 174 pages). Schweder and Hjort’s Cam- bridge University Press 2016 book Confidence, Likelihood, Probability Gruppenbild mit Dame, from one of our also benefitted from the workshop. workshops at Teknologihuset.

21 WORKSHOPS AND THE VVV CONFERENCE

May 2016: FICology ›› Focused inference ›› Model averaging Focused Information Criteria and Related Themes ›› Wars and whales

3 25 19 8 days participants talks countries

Inspiring & inspired participants, at one of our workshops at Teknologihuset, Bislett: Charkhi, Argyrou, Pircalebelu, Grünwald, Claeskens, Hegre, van Ommen, Ghosh, Gandy, Hobæk Haff, J. Vestby, Walker, Nygård, Grønneberg, Jullum, Schweder, Hjort, Stoltenberg, Østbye, Herman- sen, Moss, Van Keilegom, Cunen, Hellton. L. Walløe and M. Zolghadr also took part.

IN MAY 2016 we did FICology, on SPEAKERS WERE A. CHARKHI, G. Focused Information Criteria and Claeskens, C. Cunen, A. Gandy, P. Related Themes. The Focused Infor- Grünwald, S. Grønneberg, H. Hegre, mation Criterion was first invented K.H. Hellton, G.H. Hermansen, N.L. and developed by G. Claeskens and Hjort, I. Hobæk Haff, M. Jullum, V. N.L. Hjort, in two JASA papers Ko, T. van Ommen, E. Pircalabelu, 2003, followed by discussion contri- I. Van Keilegom, S.-E. Walker, L. butions, etc., leading also up to their Walløe, P. Østbye. In addition to Cambridge University Press 2008 many methodological issues, related book Model Selection and Model to constructing new FIC versions for Averaging. The FocuStat group has new purposes, and examining perfor- contributed to various extensions mance, two important applied talks and generalisations of the FIC, were given by H. Hegre, on forecast- including in particular efforts of C. ing incidence of internal armed con- Cunen, G.H. Hermansen, Hjort, M. flicts, and L. Walløe, on the decline in Jullum, V. Ko, S.-E. Walker, and energy storage in minke whales. these endeavours, along with those of several others, were presented at OUR CULTURAL EXCURSION this the FICology workshop. time was at the Akershus Fortress, and Schützenglene (with Nils) gave us Byrd, Byrd, Byrd.

22 WORKSHOPS AND THE VVV CONFERENCE

May 2017: Building Bridges at Bislett Bridging Parametrics and Nonparametrics

3 24 20 6 days participants talks countries

›› Bridging two worlds ›› Local parametric modelling of nonparametrics ›› Parametric models with nonparametric envelopes

THEN IN MAY 2017 we did Build- Ko illustrated novel ways of using ing Bridges at Bislett, Bridging FIC for comparing parametrics Parametrics and Nonparametrics. with nonparametrics. Also, Petrone Topics worked with included the and Prünster delved into Bayesian many ways of estimating a den- nonparametrics, with the fruitful sity or a regression function by idea of building nonparametric combining parametric models and envelopes around parametric mod- nonparametric schemes, such as els. Egeland, Rohrbeck, Stoltenberg methods generalising or further used their methodology to go into combining Hjort-Glad or Hjort- substantial applications. Jones approaches, minimum diver- gence, etc. OUR EXCURSION BRIDGED us to a guided tour of the Astronomical THOSE GIVING TALKS were R. de Observatory building from 1833, the Bin, C. Cunen, T. Egeland, I. Glad, oldest building of the University of K.H. Hellton, G.H. Hermansen, Oslo, where the congregation was N.L. Hjort, M. Jullum, V. Ko, J. also properly morley’d by Schütz­ Moss, S. Petrone, I. Prünster, C. englene (with Nils). Rohrbeck, E.Aa. Stoltenberg, B. Støve, D. Tjøstheim, S.-E. Walker. Cunen investigated ways of combin- ing parametric and nonparametric confidence distributions; Hjort discussed the Hybrid Likelihood approach involving parametric and empirical likelihoods; Jullum and

23 WORKSHOPS AND THE VVV CONFERENCE

May 2018: Vårens Vakreste Variabler Summing-up Conference

4 44 23 9 days participants talks countries

›› Summing up ›› Mozart and TTT ›› Cool strong applications

WITH THREE SUCCESSFUL based medicine for cancer treatments theme-based three-day workshops at (Frigessi), wars (Nygård), and whales Teknologihuset Bislett behind us, we (Cunen). Two journalists eagerly went for a somewhat Bigger Occa- followed the proceedings. sion in May 2018, with a four-day broader Vårens Vakreste Variabler IN THE WORDS of one seasoned summing-up conference, with about participant, which we dare to include 45 participants at the Ingeniørenes here, “I have been in academic life hus. Hjort used the occasion and since the mid-1960s and attended the Birkeland-Eyde lecture hall to many conferences, workshops Professor, Hjort, gründer, Cunen. Let's avoid nederlag for historiefaget. tentatively press upon us all the Four and the like. I have never … and, I Rules of Statistics Communication. repeat, never … had such a fulfill- ing experience. FocuStat offered SPEAKERS WERE R. DE BIN, Ø. a unique opportunity to learn Borgan, G. Claeskens, C. Cunen, I. from others. You assembled a very Eckley, H. Gjessing, J. Hannig, K.H. engaging and stimulating group. Hellton, G.H. Hermansen, N.L. Hjort, This has been a week of great I. Hobæk Haff, A. Hubin, M. Jullum, memories, including the surprise V. Ko, P. Mykland, H. Nygård, S. and delightful musical interludes. Richardson (half of the UiO’s honor- [...] Abundant future research and ary doctorates in the field of statis- practical advances will surely flow tics), L.T.S. Rønneberg, T. Schweder, from this conference.” E.Aa. Stoltenberg, A. Whitmore. Day four of the conference was set THESE MUSICAL INTERLUDES aside for Cool Strong Application gave us both Mozart’s G major KV 423 Stories, featuring long lives (Borgan), and a fiery Friday morning dixieland- personalised model- and simulation- ish TTT (for trommer, tuba, trompet). Taksdal, Kronberger, Mozart.

24 WORKSHOPS AND THE VVV CONFERENCE

Arnoldo Frigessi about new statistical approaches for VVV dinner: expressing gratitude. personalised cancer therapies.

25 2014: The FocuStat Empirical Research Kitchens Likelihood

A research kitchen is a term invented by the Steering Group of the European Science IN SEPTEMBER 2014 we had a Foundation funded project Highly Structured small meeting on Empirical Likeli- Stochastic Systems (1995–2002). The point hood, with I. McKeague and I. Van Keilegom, and some of the recently is to provide the minimum of funding and appointed FocuStat members. organisational necessities to allow a small group Hjort, McKeague, Van Keilegom of scholars, perhaps less than a dozen, to meet have worked together on earlier for a few days, to discuss and interact (and to projects (here’s a note in the IMS have meals together), without the Bulletin, 2009, on how their work on formalities of conference talks, etc. extending the scope on the empir- ical likelihood came about). This kitchen helped push ideas concern- We have had as many as five such smaller- ing how to meld ordinary parametric scale, theme-based, three-day FocuStat likelihoods and the nonparametric Research Kitchens. They have been fun, fruitful, empirical likelihood, later on leading to another joint paper. We also dis- productive, inspiring, just as kitchen work ought cussed other research ideas. to be. Our list of Kitchens is as follows.

2015: Minimum Divergence and Scoring Rules

IN SEPTEMBER 2015 we had a kitchen on Minimum Divergence and Scoring Rules, partly intended to help push S.-E. Walker’s PhD project. Our kitchen guests were F. Krüger (Heidelberg), M. Musio (Cagliari), T. Thorarinsdottir (Norwegian Computing Centre). We picked up theory for Hyvärinen scores along the way, and worked out certain generalisations and even a HIC (a Hyvärinen-based Information Criterion). Some of the minimum disparity methods we dis- cussed are intricate, causing Hjort In the kitchen spirit of things, an impromptu discussion, along with a to give a presentation on Minimum real-time construction of a HIC (a Hyvärinen Information Criterion), with Hjort’s hand, during the FocuStat Kitchen 2015. Dispair, Maximum Despair. FOCUSTAT KITCHENS

2016: 2017: L^\eta, Многая лета From Processes to Models

IN OCTOBER 2016 we did L^\eta, Многая лета, concerned with the many ways in which “the likelihood raised to a power parameter” turns up, in robustification, in Bayesian learning, in regularised estimation, etc. Guests THEN IN NOVEMBER 2017 we included T. Broderick (MIT), P. Grünwald (Amsterdam), J. Miller (Har- did From Processes to Models, on a vard), P. Müller (Austin, Texas), A Ghosh, and locals I. Glad, P. Mykland. list of themes related to the concept of building process models “behind the data”, in various connections and applications. The general idea is that when such background models are biologically or context-wise plausible, they lead to better models for the data themselves, compared to off-the-shelf type methods for these. All FocuStat members were on board, and guests A happy group, learning more about the many ways in which we may raise the likelihood were R. de Bin, S. Engebregtsen, C. to a power parameter: S.-E. Walker, G.H. Hermansen, K.H. Hellton, I.K. Glad, N.L. Hjort, J. Miller, P. Grünwald, T. Broderick, C. Cunen, R. de Bin, P. Mykland, E.Aa. Stoltenberg. J. Heinrich, B. Lindqvist, J. Moss, R. Par- Argyrou, A. Ghost, P. Müller also took part. viero, T. Thorarinsdottir, M. Tveten.

2018: Combo Kitchen The Combo Kitchen!, on Combining November 2018. Diverse Information Sources

FINALLY WE HAD our Combo Kitchen in November 2018, con- cerned with methods for combining statistical information across very diverse sources. In addition to the core group, and several Oslo locals, we had I. Carmichael (Chapel Hill), D. De Angelis and A. Presanis (Biostatistics Unit, Cambridge), M. van Nee and M. van der Viel (VU, Amsterdam). Themes included the II-CC-FF meth- ods of Cunen and Hjort, evidence synthesis, empirical Bayes, Bayesian co-data learning in ridge models, and drug interaction prediction, and The Combo Kitchen crew, eagerly learning to combine information and knowledge applications to government directed sources. The music stand was used for the Telemann duet. From left: V. Vitelli, C. Cunen, vaccination policies in the UK. A. Presanis, I. Helland, R. Parviero, J.F. Schenkel, D. De Angelis, N.L. Hjort, M. Zucknick, M. v Nee, I. Carmichael, M. vd Wiel, K.H. Hellton.

27 DURING THE PROJECT we’ve organised Important Evaluation VIPs, Guests, Weeks, featuring first Alan Gel- fand, Duke University (November Collaborators, 2015), and then Per Mykland, University of Chicago (November Networking 2016), as our one-week evaluators. This meant presenting parts of our projects to these Very Important In addition to being active on the work- Professors, theme after theme, result after result, application after shop and conference fronts, we have also applications, where the VIPs then had a fair stream of guests, for shorter and gave us important feedback, from ­longer time, for exchanges of ideas, and for constructive criticism to a bit of ­serious collaborative work. Some of us have praise. It takes time and forces, for established long-term working relationships all concerned, but is a very fruitful with colleagues, from Oslo and abroad. way to learn, including learning what is still remaining for a piece of work to be good enough, or to become very good. Also important have been our friendly but con- structive local critics, professors Arnoldo Frigessi and Ingrid Glad, who have followed our progresses and processes and offered advice and wisdom.

PROFESSOR EMERITUS WITH UiO, Tore Schweder, deserves spe- cial mention in these paragraphs, since he has been a positive and constructive friend of the full group, and for his long-term productive working relationship with Hjort. They finally managed to finish their magnum opus, the Confidence, Likelihood, Probability (500 pages with the Cambridge University Press, 2016), and were guest editors for the Special Issue on Confidence

Hard at work, during the Mykland Week, November 2016: The Schweder and Hjort Confidence, Hjort, Hermansen, Stoltenberg, Hellton, Cunen, Mykland, Likelihood, Probability 500-page book is Walker. Ko took part for some of the days. There are out; here at the ISBA in Sardegna, June traces of cleverly invented weighted log-likelihoods on the 2016, with our Cambridge University blackboard. Press agent Diana Gilloly. VIPS, GUESTS, COLLABORATORS, NETWORKING

Distributions and Related Themes, to us taking active part in the Sci- of the Vietnam War and Flower with Journal of Statistical Planning entific Committee meeting of the Power, as a likely change-point, and Inference. This Special Issue IWC, in Bled, Slovenia, May 2017. with the battle-death distribution came about partly as a result of the Here Cunen presented several for great wars changing for the successful FocuStat workshop Infer- analyses, and had to stand firm in better then. Steven Pinker’s quick ence With Confidence. a stream of countering views from twitter reponse, “Sophisticated new Australian researchers, almost in a analysis of the stats of war by Nils IN THE LIVES of academics and courtroom situation. Notably, Hjort Lid Hjort affirms a decline over research one rarely clings to One has also worked with Walløe and time, but finds the sharpest in 1967, Plan Only, as other matters tend to his colleague Marianne Thoresen not 1945”, spurred further atten- cross our desks, or turn up as ques- regarding a partly controversial tion and interest in these statistical tions to pursue during workshops journal article on the cooling of models and analyses, and has led or via personal contacts, whether newborns after oxygen deprivation to the Statistical Sightings of Better by inci- or accident or not. Two during birth; see Hjort’s blog post Angels paper with Cunen, Hjort, important strands of Unplanned for the issues under debate. Nygård. Also, G.H. Hermansen is Work for us, in particular for C. pursuing joint work with Nygård, on Cunen and N.L. Hjort, have turned the escalation of conflicts. There will out to involve whales and wars. be more joint work, with some of These have escalated into Serious us and some from PRIO, including Research and the time-consum- efforts to develop and finesse sta- ing business of writing scientific tistical methods for change-points. papers, giving presentations at Walløe, Cunen, Hjort at the Scientific What are the mechanism securing conferences, and even being inter- Committee meeting with the International stability, and which underlying viewed by media. Also G.H. Her- Whaling Commission, Bled, Slovenia, parameters and processes might 2017, complete with pink whales on our mansen is involved with the wars. name tags, apparently. lead to conflict?

THESE ENDEAVOURS HAVE WE WISH TO convey that our crew come about via personal contacts have maintained a group spirit with Lars Walløe (senior scientist in throughout our project, also when it multiple areas, chemistry, physiol- comes to new ideas coming in and ogy, medicine, statistics, advisor to floating around. When the whales the government, former President surfaced in front of us and the wars Nygård, Hjort, with master student J.K. Haug of Academica Europeae, etc.) and (jointly supervised by Hjort and Nygård), popped up, those of us who pursued Håvard Nygård (senior researcher going into war and peace research. Hjort the themes actively also benefitted and research director with Peace reads Война и мир, the others read Pinker’s from weekly discussion meetings Better Angels of Our Nature. Research Institute of Oslo). and insights emanating from these. PRIO ORGANISES A series of WALLØE HAS BEEN involved with Oslo Lectures on Peace and Conflict, IN ADDITION TO Schweder, the Scientific Committee work of and in January 2018 Hjort was Walløe, Nygård, other scholars are the International Whaling Com- invited to be a panel discussant after on our collective interactive map mission, for a number of years, and a lecture by Aaron Clauset, where of active collaboration. Hellton one specific research theme has the claim was made that the world has worked with and continues to raised considerable interest and of wars has remained essentially work with Claudio Heinrich, Lasse controversy: Have the Antarctic stationary over the past 200 years. Giil, Thordis Thorarinsdottir, Jo minke whales become thinner, over Hjort dared to disagree with this, Røislien; Hermansen with Steffen the past few decades, or not? And however, finding significant traces Grønneberg at BI and as mentioned what is the role of statistical model of evidence in the benevolent with Nygård; Hjort also with Gerda selection here? Initial discussions direction of “the world is slightly Claeskens, Ingrid Glad, Ingrid with Hjort, and then Cunen, led to more peaceful now”. In the Towards Hobæk Haff, Ingrid Van Keilegom; research work and new analyses, in a More Peaceful World [insert `!’ Ko with Fabian Gieske; Stoltenberg the form of journal papers to come, or `?’ here] blog post, written up with Per Mykland and his co-­ lectures and seminars, interviews after this event, arguments are also supervisor Sven Ove Samuelsen; and by some media, and also, crucially, presented pointing to 1967, the time Walker with colleagues at NILU.

29 Excursions: FocuStat Workshops and Excursions Conferences

We have been unusually busy travelling to IN AUGUST 2015 we went to exotic various workshops and conferences, along Rio de Janeiro, for the 60th Inter- with other research visits, though not national Statistical Institute Con- all of us all of the time. ference. Hjort gave an invited talk on combining information sources via confidence distributions, Cunen We decided early on, however, that we for talked about Poisson models for each year should single out one conference, two-by-two tables, Hellton about where we all tried to go (and to give lots high-dimensional principal com- of talks, and otherwise being active). In ponents, Hermansen on FIC for addition there has been a list of cultural time series, and Walker on proba- bilistic forecasting of air pollution. excursions on our agenda. Incidentally, before Rio most of us also went to the Norwegian Statistical Meeting in Solstrand, giving a significant percentage of the presentations. In June 2016 we took eager and high-visibly part in the NordStat 2016, the Nordic Meeting on Mathematical Statistics, in København. Cunen and Hellton won Best Poster Awards, and Hjort had an invited session, featuring also the hybrid likelihood.

AFTER RIO AND København we upped the ante in June 2017 by going to the Norwegian Statistical Meeting at Fredrikstad, again giving lots o’ talks. And in June 2018 we again went to the NordStat, where this particular Nordic Meeting on Mathematical Statistics took place in Tartu, Estonia. Cunen had the strenuous privilege of organising an invited session, on confidence distributions, and the others gave talks in different sessions.

AMONG THE MANY other work- shops and conferences we’ve taken part in, for the years 2015-2018, and where several of us have pre- In Rio, checking out the latest in professorial sented invited lectures, we single papelaria paraphernalia: Cunen, Hermansen, Walker, Hjort, Hellton. out the following. FOCUSTAT EXCURSIONS

Clinical Biostatistics, Wien. Then in young students, etc. She gave in 2015, Hjort gave such invited talks particular an invited talk at the 61st at the 60th International Statistical International Statistical Institute Institute World Congress in Rio World Conference in Marrakech, de Janeiro, and E.Aa. Stoltenberg and gave several presentations at the represented Norwegian statistics at Scientific Committee Meeting of the the 19th European Young Statisti- International Whaling Commission, cians Meeting in Prag. Invited talks in Bled. Hjort gave six hours of invited occasions for Hjort in 2016 included lecturing at the annual Geilo Winter Various core and associated FocuStat the Fusion Learning and BFF Con- School for Applied Mathematics. He members in historical Tartu, June 2018: ference at Rutgers, the Joint Sta- had other invited talks, including at K.F. Frøslie, E.Aa. Stoltenberg, C. Cunen, N.L. Hjort (back row); J. Moss, S.-E. tistical Meeting in Chicago, and the the BFF in Harvard, the Bayesian Walker, G.H. Hermansen, V. Ko (front row). Post-Selection Workshop at Leuven; Nonparametrics Conference in Paris, also, he was invited session organ- and the Joint Statistical Meeting in IN 2013, N.L. HJORT won the iser at the International Society for Baltimore. Cunen, Hjort, Stoltenberg Sverdrup Prize and gave a talk at Bayesian Analysis in Sardegna. K.H. were invited speakers at the Sylvia the prize award occasion; he was Hellton was an invited speaker at an Richardson Honoris Causa workshop also invited to give the Corcoran Oberwolfach Workshop on high-­ in Oslo. Then in 2018 both Cunen and Memorial Talk at the Department of dimensional data. Hjort appeared on certain panels, at Statistics, University of Oxford. In peace-and-conflict arrangements at 2014, Hjort gave invited talks at the 2017 MARKS THE year where C. the Peace Research Institute of Oslo, 2nd World Conference on Non- Cunen in particular starts getting a and Cunen was invited to organise a parametric Statistics, Cadiz, and at high number of invitations to give three-lecture session at the NordStat the 35th International Society for talks, take part in panels, speak to Meeting in Tartu.

Excursions: all the way to Kobberhaugen; and we’ve read Houellebecq’s Soumis- Cultural sion for a book discussion evening. Once upon a time we even went Excursions to the cinema, to watch “Focus” (which we didn’t quite understand).

THERE HAS ALSO been various THE FOCUSTAT CREW have musical pre-, inter- and postludes Blindern Stunt- og Geriljakor singing after Cunen’s PhD disputation, about to tallis also taken part in various cultural at our kitchens, workshops, confer- the audience. excursions (with and without gentle ences, and meetings. The Schütz- pushing). We’ve sang four-part englene (with Nils, and led by Tor duet). When the Erling Sverdrups specially made songs for several Tveite) have been singing, appar- Plass was offically opened, at the of our end-of-semester meetings; ently specialising in the 1590ies 8th floor in the Abel Building, in once we composed ten verses of repertoire (but with different com- September 2015, our cellists Cunen FocuStat themes with an arrange- posers and languages, from work- and Hellton played Kummer. ment of Revolusjonens røst, for four shop to workshop); Nora Taksdal voices and two celli; we’ve been to and Janina Kronberger have played IT WAS WITHIN the crew spirit of the University Symfoniorkester to the KV 423; Are Rosseland, Mar- things, then, that Nils hired Blind- hear Brahms (with Cunen on the ius Gjersø, Steffen Granly cleverly ern Stunt- og Geriljakor, led by Tor cello); played pieces for two celli interrupted the VVV proceedings by Tveite, to sing about wars, whales, and piano; we’ve watched Emil their TTT (trommer, trompet, tuba); and angels, after Cunen’s PhD Stoltenberg fly through the air (with and in the Combo Kitchen we had disputation (where her dissertation Nordberg Basket winning against Frederik Vogelfänger Aakre (with themes included, indeed, wars, Persbråten); we had a skiing day Nils joining him to play a Telemann whales, and angels).

31 THE FOCUSTAT BLOG has proven The FocuStat to be a success story, in terms of numbers of visitors (some of Blog our blog posts have been read by thousands), and by inspiring further attention and discussion (some have The FocuStat crew have actively used led to interviews and media atten- The FocuStat Blog as an outlet for tion). It often takes more than half telling Statistical Stories, related to a year to hammer out a solid article for a top journal, even after the main matters that cross our desks or ideas have been understood and otherwise catch our attention. worked out, and another half year to have it accepted and published. The traditional outlets for conveying But when N.L. Hjort used one long weekend to write up a blog story on research, for academic and applied whether the world has become more statisticians, remain those of journal peaceful or not, based on war battle articles and talks at seminars and death statistics from 1823 to the conferences, whereas many stories present, it was read by thousands deserve to be told in a crisper fashion after a few weeks (partly thanks and to a broader audience, regarding to Steven Pinker, who tweeted both the methods themselves and their praise for the analysis). Similarly, C. Cunen’s blog post comparing applications. Game of Thrones with the Wars of the Roses managed to attract broad international attention.

OTHER BLOG HITS have been those concerning sports – statisti- cal stories related to speedskating, cross-country skiing, football, handball, have been picked up by many. After one such blog post, Hjort took part in Abels Tårn, explaining to the people why the Olympic 1000 m sprinters should skate twice. Also, Norway won more medals than any other nation in the 2018 Olympics, but what does a proper binomial counting process analysis say?

THE FOCUSTAT BLOG shall go on, with new Statistical Stories, also after the formal end of the FocuStat project. As of January 2019, a partial listing of the Blog Posts is as follows.

Towards a More Peaceful World [Insert ‘!’ or ‘?’ here] is one of many Blog Posts which have caught international attention. THE FOCUSTAT BLOG

Céline Cunen (left) is becoming world famous, apparently. Her FocuStat Blog Post Mortality and Nobility in the Wars of the Roses and Game of Thrones (2015) has attracted world-wide attention. She has given several talks and interviews about the relevant statistical, historical, science-fic- tional themes, and different types of media, in at least a dozen different languages, have discussed her work.

N.L. Hjort: N.L. Hjort: once an innocent looking rule is dec oct Norway-Spain 28-25: Conspiracy Probability changed. 2014 2015 How exciting was it? Calculus for Norwegian Footballers Poisson models are used to com- pute and display Real Time Real When the Norwegian Football N.L. Hjort: mar Excitement Plots™, the probabil- Association carried out a so-called Recruitment Dynamics 2016 ities of team A winning, or team B random drawing for the 8 matches and Stock Variability: winning, or having a draw, in real in the round-of-16-teams, towards The Johan Hjort Symposium, some time, as the match proceeds. the end of the Cup Tournament, the personal reflections result was met with disbelief. Was a conspiracy under way? This post was written up by the invitation of the Hjort Centre K. Hellton: of Marine Biology, giving the sep Hvor stemmes det hva i writer another chance to think 2015 hovedstaden? M. Jullum, D.R. Baños, through the life and work of his oct N.L. Hjort: great-grandfather. 2015 Local differences and variation in To liv: kvinnene i Oslo people’s voting patterns are Lillestrøm som ble født på samme examined. dag og døde på samme dag C. Cunen, N.L. Hjort: jan Two girls are born on the 14th of New statistical methods 2017 December, 1912, and 101 years and shed light on medieval C. Cunen: 11 months later they die at the same literary mystery (January 2017). oct Mortality and Nobility in day, in Lillestrøm. How can we give 2015 the Wars of the Roses reasonable answers to questions Tirant lo Blanc is arguably the and Game of Thrones like “how unlikely is this”? world’s first novel, written in Catalan in the 1460ies. When did Analyses of survival time, including Author Two take over for Author the probability of having a violent One? death, for different strata, are carried K. Hellton: dec out, and then compared, for two The bonus roll and the 2016 comparable universes. Christmas spirit

A popular “Christmas roll” changes its winning chances dramatically

33 THE FOCUSTAT BLOG

N.L. Hjort: T. Schweder: N.L. Hjort: jan nov feb But some are more equal Bayesian Analysis: One Thousand is Unfair, 2017 2017 2018 than others. Always and Everywhere? Two Thousand is Fair

Almost miraculously, Allan Dahl As our guest writer, Schweder The 500 m and 1000 m events are Johansson and Odin By Farstad spells out why even a Nobel Prize the Formula One challenges of ended up having exactly the same winner may be wrong, through speedskating, and their the intrica- point-sum after four distances of unforeseen consequences of a cer- cies of inner and outer lanes matter, speedskating in the Norwegian tain type of a Bayesian prior. The in an Olympically significant fash- Junior Championships. Just how confidence distribution approach ion. Medals may change necks, if a phenomenally rare is this? works rather better. skater is given another start lane.

N.L. Hjort: N.L. Hjort: N.L. Hjort: feb des feb The semifinals factor for Cooling of Newborns The Best Metal-Grab- 2017 2017 2018 skiing fast in the finals and the Difference bing Games Ever Between 0.244 and 0.278 Petter Northug won a bronze in the Norway had an astounding metal ski sprint event at the 2010 Vancou- Cooling of neonates can save small count, at the 2018 Olympics, but ver Olympics – but why not the gold? lives, after traumatic births with surely a binomial process approach oxygen deprivation. This depends to medal counting is more appro- crucially on the time window, how- priate than mere counting when ever, and here a recent JAMA 2017 factoring in that fewer events and N.L. Hjort: study is criticised. fewer nations took part in earlier feb Ut på tur: NordStat 2016, Olympics. 2017 København juni 2016

A report is given from the FocuStat C. Cunen, N.L. Hjort: jan excursion to NordStat 2016, where we Whales, Politics, and N.L. Hjort: 2018 jul took eager part, and won two prizes. Statisticians. Belgium Breaks a 48 2018 Year Old Curse (Appar- Just occasionally our research ently). enters the hot soup of controversy E.Aa. Stoltenberg: and politics and international cri- The media claimed it was almost on jul Frekventisme, Bayes og cism, as here, where we presented a miracle level, when Belgium turned 2017 likelihoodprinsippet our analyses for the slimming-or- 0-2 against Japan to a 3-2 victory. not of the Antarctic Minke whales, But how improbable was this? This is a philosophically inspired for the Scientific Committee of the essay on the two-or-more schools International Whaling Commission. of statistics. N.L. Hjort: aug Overdispersed 2018 N.L. Hjort: Children jan C. Cunen, G.H. Her- Towards a More nov 2018 mansen, E.Aa. Stolten- Peaceful World [insert The boy-to-girl ratio is not quite 2017 berg: ‘!’ or ‘?’ here] 50-50, and neither is the girl-prob- Bokmelding: Confidence, Likeli- ability constant across families, hood, Probability (Schweder og We analyse the horrible sequence and perhaps not even inside Hjort, CUP 2016) of battle death counts, in all major families. It turns out children are interstate wars since 1823, and both (slightly) overdispersed and Three eager readers give their point to statistical evidence of (slightly) Markovian. account of the CLP book. better angels.

34 STATISTICAL STORIES

Xxx Awards & Prizes

Somewhat rarely but then always jubilantly, prizes & awards are being bestowed upon FocuStat members. Here’s our list.

K.H. Hellton and C. Cunen won the Best Poster Presentation Awards at the NordStat Conference, København, June 2016, here in front of Cunen’s poster.

C. Cunen: Best Poster Presentation, the 25th year of Department of Sta- University of Oslo (supervisor: N.L. NordStat 2016, København. tistics, University of Oxford, 2013. Hjort), November 2013.

C. Cunen: The Titan Prize 2018, for N.L. Hjort: elected member of the M. Jullum: Ahtlete of the Month, science communication. Kongelige Norske Videnskabers for the International Orienteering Selskab (the Royal Norwegian Society Association, August 2014. C. Cunen, representing Norway, of Sciences and Letters), January European Young Meeting of Statis- 2017. He is also elected member E.Aa. Stoltenberg, representing ticians, Beograd, 2019. of the Videnskaps-Akademiet (the Norway, European Meeting of Norwegian Academy of Science and Statisticians, Prag, 2015. K.H. Hellton: Best Poster Presenta- Letters), since 1999. tion, NordStat 2016, København. E.Aa. Stoltenberg: Fulbright M. Jullum: The Norwegian stipend, for working at the Steva- G.H. Hermansen (with N. Foldnes Computing Centre Prize, for best novich Center for Financial Mathe- and S. Grønneberg): The Cappelen Master Thesis within mathematics matics at the University of Chicago, Damm faglitterært forfatterstipend and information sciences, at the August 2018 to June 2019. (NOK 100.000), for their book writ- ing project, 2015. The resulting book Statistikk og dataanlyse: En moderne innføring was published in 2017.

N.L. Hjort: Sverdrup Prize, 2013. This prize is awarded every second year, by the Norwegian Statistical Association, to an eminent repre- sentative of statistics in Norway. Its first recipients have been Dag Tjøstheim (2009), Tore Schweder (2011), and now Hjort (2013). Later winners are Odd Aalen (2015) and Ørnulf Borgan (2017).

N.L. Hjort: giving the Corcoran C. Cunen won the Titan Prize for science communication 2018. Here she is flanked by Titan’s main editor Wenche Willoch and Vice Dean Kristin Vinje wth the Faculty of Mathematical Memorial Lecture, at the occasion of Sciences, University of Oslo.

35 FocuStat in 2014: Our PhD student Martin jul the Media Jullum is World 2014 Champion in Pre-Orien- teering, in Trentino, Italy. FocuStat projects and participants occasionally attract the attention of the Our World Champion aug Martin Jullum is Athlete media. Here’s a partial list. 2014 of the Month of the International Orienteering Federation.

General release occa- des sion, with a bit of media 2014 attention, of the Ad Fontes Festskrift for Terje Kvam, where Nils Lid Hjort has contrib- uted his Deilig er Jorden essay.

2015: Denne 27-åringens apr regnestykker skal gi 2015 bedre kreftbehandling, claims Budstikka, about Kristoffer Hellton.

Martin Jullum wins the jun bronze in the Pre-Orien- 2015 teering World Champi- onship in Zagreb, Croatia.

Matematikere på oct Blindern har regnet ut 2015 sannsynligheten for sammentreffet mellom Borghild og Khuu. «Hvis man sjekker en billion kvinner, vil man finne fem slike par», explains Martin Jullum and David Ruiz Baños, to Aftenposten- TV.

Aftenposten’s oct A-Magasinet has a 2015 special edition about “Two Lives”, with the fascinating life stories of two women, born on the same day, who died on the same G.H. Hermansen, C. Cunen, N.L. Hjort were featured in forskning.no day, in Lillestrøm, 101 years later. (December 2018) and in Apollon (January 2019), in connection with their statistical analyses of war-and-conflict processes. They asked for help regarding the FOCUSTATSTATISTICAL IN THE STORIES MEDIA

Xxx K. Hellton in Budstikka, explaining high-di- mensional regression models for person- alised predictions.

simple-but-complicated question Céline Cunen is featured Emil Aas Stoltenberg is sep nov “how unlikely is this”. Martin in Universitas: The in Dax18, Dagsnytt 2016 2016 Jullum, David Ruiz Baños and Nils death rate in Game of Atten, NRK, to explain Lid Hjort are credited in the special Thrones is closer to real history US elections to the people; he edition for assessing and present- than you think. returns the next day, also in Dax18, ing different aspects of that to explain more. question. This also led to the «Bra påvist! Påvis oct FocuStat Blog Post To liv: kvinnene mer!», writes selveste UiO-forskere 2016 dec i Lillestrøm som ble født på samme Natt og Dag, about undersøker om Donald 2016 dag og døde på samme dag. Céline Cunen’s work on the Wars Trump er julenissen, of the Roses. reports Titan, again relying on the Centre of Apparent Excellence for Vinnie Ko has his Jule- og Julenisse-forskning, where nov splendid book Met N.L. Hjort seemingly is on board. 2016 2016: Hartelijke Groente Gudmund Hermansen launched in the Netherlands. may behind the success of 2016 the Oslo Data-Science November 2016, and for nov Meetup, explains Big Insight. a long row of later 2017: 2016 occasions: Vinnie Ko Céline Cunen is featured jan Martin Jullum grabs the has so many interviews and in Titan: New statistical aug 2017 silver in the Pre-Orien- rescensions and book signing methods shed light on 2016 teering World Champi- occasions that he could easily have mediaeval literary mystery. onship (after Martin Fredholm) in his own “in the media” list – and, Strömstad, Sweden. inderdaad, he has; check his in de Céline Cunen is giving apr media here. an «åpen dag» Univer- 2017 Céline Cunen is featured sity of Oslo lecture, also aug in Titan: The death rate Emil Aas Stoltenberg shown (and kept) on television, 2016 nov in Game of Thrones is takes part in Urix, NRK, NRK Skole. 2016 realistic. concerning the ins & outs of the US elections.

37 FOCUSTAT IN THE MEDIA

Arbeiderpartiet faller, aug men vinner: Finans­ 2017 avisen interviews 2018: Kristoffer Hellton, Anders Løland Céline Cunen and Nils jan and Clara-Cecilie Günther. Lid Hjort are portrayed 2018 in Titan: Mest sannsyn- Aber Ihre Untersuchung lig at Den lange freden begynte sep war nicht nur Spaß, under Vietnamkrigen. 2017 Céline Cunen?, fragt Technology Review, in their Steven Pinker tweets jan interview, for readers in three admiration for Nils Lid 2018 countries: Game of Thrones ist Hjort’s blog post historisch realistisch. Towards a More Peaceful World [Insert `!’ or `?’ here]. Ender med å velge sep det trygge: Dagens Nils Lid Hjort is a panel 2017 jan næringsliv interviews discussant for the PRIO 2018 Kristoffer Hellton. event Oslo Lectures on Peace and Conflict. Jo da, miljøfartsgrensen nov virker. Article in January 2018: Nils Lid 2017 jan Aftenposten, about air Hjort and Håvard 2018 quality in the Oslo region, by Britt Nygård are portrayed in N.L. Hjort is properly starstruck when tak- Ann Kåstad Høiskaar (NILU), Dag Titan: Uvisst om Den lange freden ing part in Abels Tårn (February 2018). Tønnessen (NILU), Sam-Erik er kommet for å bli. Walker (NILU and the FocuStat the War-and-Peace FocuStat Blog group). Nils Lid Hjort, now qua Post, and was starstrøkk again when jan literary analyst, on the Abels Tårn panel. The 2018 The Centre of Excel- apparently, makes it to background, for both the Abels Tårn dec lence for Jule- and the =Oslo. appearance and the Titan article, is 2017 Julenisse-forskning, the Pre-Olympic 2 x 1000 meter with Hjort among the alleged Om å spise opp en rygg. FocuStat Blog Post. feb experts, is portrayed in Titan: Du Nils Lid Hjort on NRK 2018 kravlende, glitrende, krypende, P2 radio, the cultural «Kvinner er forskjellige feb kravlende ... channel (of course), 22-ii-2018. fra menn, rent aerody- 2018 namisk sett, forklarer Another entry in Titan, Last ned statistikeren professor Hjort.» Skøyteløp: Start i dec feb with the same Centre of som endra OL (og vil indre bane gir stor fordel på 2017 2018 Excellence for Jule- and gjøre det igjen). Podcast, 1000-meteren. In Titan. Julenisse-forskning: Har julenissen from Abels Tårn (Vaffel og spist for meget rød fluesopp. Vitenskap), 16-ii-2018; tv pro- Skøyteløp i OL blir aldri feb gramme version aired on 20-ii- helt rettferdig. Says Nils 2018 2018, NRK2. Nils Lid Hjort was Lid Hjort, apparently, in starstrøkk some weeks earlier, when forskning.no, 23-ii-2018. Pinker tweeted his admiration for

38 FOCUSTAT IN THE MEDIA

PRIO reblogs Nils Lid The Long Peace most The Verden er blitt feb jun dec Hjort’s blog post, with likely began during the fredeligere story has 2018 2018 2018 Towards a More Vietnam War. Titan been picked up by Peaceful World [insert `!’ or `?’ [translation of earlier feature forskning.no, and is front page here] filed under “Conflict Pat- article, from Norwegian to news there, the first week of terns” (and tagged with “Better English]. December. Angels” and “Vietnam War”). Sørlig vågehval er December 2018: Verden sep dec Jeg elsker dig! © Nils avmagret. In Klasse­ er blitt fredeligere (the apr 2018 2018 Lid Hjort – each year, a kampen (yes, Class world has become more 2018 few of Nils’s translations Struggle, 21-ix-18), about the peaceful). Article in Apollon, of classical song texts find their eleven years of statistical-political featuring Céline Cunen, Gudmund ways to official concert programme fighting within the IWC, now Hermansen, Nils Lid Hjort, Håvard notes; the most popular remains “won” by PhD candidate Céline Nygård. See also the Apollon this Grieg song, as here, with Cunen and professor Nils Lid Theme Issue on contemporary soprano R.R. Harmann, April 29 Hjort. applied statistical research, 2018, for an evening recital with January 2019. Love Songs, here with the Ohio Endelig enighet om at sep Valley Symphony. den sørlige vågehvalen Céline Cunen is the 2018 dec er blitt avmagret. Story Titan Prize 2018 2018 Do we need May 17th in Titan, concerning how statisti- Awardee. In Titan, the may every year? «Å fjerne cians Céline Cunen and Nils Lid University of Oslo: Prisen utdeles 2018 allround på skøyter er Hjort entered the international and årlig til en forsker som har gjort en som å endre reglene i fotball for å partly politicised debate on the fremragende innsats med å legge opp til flere mål.» Kronikk, slimming of minke whales in the formidle forskningen ved VG, by Arve Hjelseth (sosiolog), Antartic, in the International fakultetet. Cunen har utmerket seg Iver Ørstavik (filosof), and Nils Lid Whaling Commission’s Scientific ved å kommunisere om statistikk Hjort (statistiker og professor, Committee (and won, somehow). på en overbevisende, morsom, UiO). viktig og interessant måte. Game of Thrones and sep Kunnskapsforetak? Nei Reality: A Cross-disci- may 2018 takk! La oss slippe å plinary start of the term, 2018 bruke tid på dette nå at PRIO, with Céline Cunen giving igjen – vi vil forske, undervise og a lecture and then being on the formidle! Kronikk, Aftenposten, panel, with 125 pizza eating 2019: 2-v-2018 (with Nils Lid Hjort participants. In Apollon's special jan among a string of other professors issue on applied 2019 about foretaksmodellen). Krig, fred, statistikk statistics, as many as dec – om at verden er blitt three stories feature FocuStat work. 2018 Two statisticians at the fredeligere. Radio jun University of Oslo have conversation in NRK2 Studio 2, 2018 blown a hole in Steven with Nils Lid Hjort and Håvard Pinker’s famous theory that the Nygård, interviewed by Otto Haug. Long Peace dates from 1945 onwards. A version of the Titan article, re-blogged at PRIO.

39 TALKS & PRESENTATIONS

Talks & Presen­ tations

Talks and presentations given by the FocuStat participants, including also those from 2013 of relevance for the 2014-2018 project.

There’s always an element of polite excitement in the hall when one of us is about to give a lecture. This is from the ISI 2013 Conference in Hong Kong, half an hour before N.L. Hjort gives his invited talk on Confidence, Likelihood, Probability.

2013: • N.L. Hjort: Analysing an ensemble of 2 x 2 tables with many zeroes. BMMS Statistics Seminar, November 2013. • N.L. Hjort: Mathematics, physics, probability theory and pedagogical principles in Duck- burg. (Exceedingly) popular science talk, 2014: Bjørnegildet, University of Oslo, February 2013. • N.L. Hjort: Statistisk dag: Confidence and • N.L. Hjort: Quiet Does Not Flow the Don: Likelihood. Tore Schweder 70 year confer- Statistical analysis of a quarrel between ence, Academy of Science and Letters, Nobel laureates. Ski og matematikk, Oslo, April 2013. Ronda­blikk høyfjellshotell, January 2014. • M. Jullum: Parametric or nonparametric: • N.L. Hjort: Stille flyter ikke Don: Statistisk The FIC approach. 17th Norwegian Statisti- analyse av en krangel mellom nobelpris- cal Meeting, Halden, June 2013. vinnere. Sinus-konferansen, Cappelen-­ Damm, Oslo, March 2014. • N.L. Hjort: Beta-tilted Beta Processes. Invited talk, Bayesian Nonparametrics IX, • N.L. Hjort: Distributions of confidence. Amsterdam, June 2013. Department seminar at Norwegian Busi- ness School, April 2014. • M. Jullum: Parametric or nonparametric: The FIC approach. 29th European Meeting • N.L. Hjort: FocuStat: Focus Driven Statis- Statisticians have worked out how to meet of Statisticians, Budapest, July 2013. tical Inference with Complex Data. Lunch interesting people – between talks at the seminar at the Norwegian Computing 60th ISI World Congress in Rio de Janeiro, • N.L. Hjort: Confidence, Likelihood, Prob- Centre, April 2014. Julty 2015. ability. Invited talk, ISI World Conference, Hong Kong, August 2013. • N.L. Hjort: HEL: A hybrid empirical likelihood bridging from nonparametrics to paramet- pionships: how the semifinals influence the • N.L. Hjort: Distributions of Confidence, rics. Invited talk at the 2nd World Confer- finals. Fagligpedagogisk dag, University of ­Norwegian Computing Centre, Oslo, ence on Nonparametric Statistics, Cadiz, Oslo, October 2014. October 2013. June 2014. • N.L. Hjort: Model comparison, model selec- • N.L. Hjort: Distributions of Confidence. • S.-E. Walker: Ensemble based probabilistic tion and model averaging. SFI^2 Farewell Invited talk, The International Selbu Confer- forecasting of meteorology and air quality in Workshop, University of Oslo, November ence, Selbu, October 2013. Oslo, Norway. Presented at WWOSC 2014: 2014. • N.L. Hjort: Distributions of Confidence. The World Weather Open Science Con- Corcoran Lecture, Department of Statistics, ference, Montréal, Canada, August 2014. • N.L. Hjort: Are we in fact faced with one Oxford University, October 2013. (Based on work carried out with NILU, the of the most flagrant cases of plagiarism in the history of literature? “Mathematical • N.L. Hjort: Who should be allowed to skate Norwegian Institute for Air Research.) pearls and curiosa” lecture, University of the 10,000 metre? Fagligpedagogisk dag, • N.L. Hjort: What price Cox regression? Trondheim, November 2014. University of Oslo, October 2013. Ranking estimates from semiparametric • N.L. Hjort: Last inner track: Does one need and parametric hazard regression models. to race the 500 metre twice? «Fra tall til inn- Invited talk at the ISCB 35, Wien, August sikt» evening with statistics, Litteraturhuset, 2014. Oslo, November 2013. • N.L. Hjort: Cross-country ski sprint cham-

40 TALKS & PRESENTATIONS

2015: • A. Øygard: Deep learning “hands-on” tuto- rial. Oslo Data Science Meetup. Teknologi- huset, Oslo, October 2015. • C. Cunen: Optimal inference via confidence • C. Cunen: Confidence distributions: a distributions for two-by-two tables: presen- general introduction and an application to tation and comparisons with Liu, Liu, Xie meta-analysis of two-by-two tables [joint methods. International FocuStat Workshop work with N.L. Hjort]. Seminar at the Biosta- “Inference With Confidence”, Oslo, May 2015. tistics group at the Norwegian University of • K.H. Hellton: PCA and the asymptotic Life Sciences, October 2015. distribution of high-dimensional sample • N.L. Hjort: Tirant lo Blanch, 1490: When eigenvectors. International FocuStat Work- did Author Two take over for Author One? shop “Inference With Confidence”, Oslo, Fagligpedagogisk dag, University of Oslo, May 2015. October 2015. • G.H. Hermansen: Predictive confidence for • N.L. Hjort & C. Cunen: The World’s First With Sonia Petrone and Igor Prünster, time series. International FocuStat Work- Novel: When did Author Two take over Bayesian Nonparametrics Bridging at shop “Inference With Confidence”, Oslo, for Author One? Pub med professor (og Bislett, May 2017. May 2015. PhD-student), Realistforeningen, Univer- • N.L. Hjort: Confidence distributions for sity of Oslo, November 2015. change-points. International FocuStat Work- shop “Inference With Confidence”, Oslo, vidual clusters. Poster & flashtalk (winning May 2015. one of three Best Poster Awards), NordStat 2016: 2016, København, June 2016. • S.-E. Walker: Confidence distributions based on M-estimators. International • G.H. Hermansen: Focused model selection FocuStat Workshop “Inference With Confi- • N.L. Hjort: Statistikk, sannsynlighetsteori, strategies for time series. Poster & flashtalk, dence”, Oslo, May 2015. sjansespill, samfunn, solidaritet. Part of the NordStat 2016, København, June 2016. ungdomsundervisningen, for 14-15 year • N.L. Hjort: The hybrid likelihood: Bridging • C. Cunen: Optimal inference for two- old school pupils, organised by Humanist- parametrics and nonparametrics. Invited talk, by-two tables: a confidence distribution foreningen at UiO, February 2016. approach. PhD and PostDoc workshop at NordStat 2016, København, June 2016. • K.H. Hellton: When are principal component Klækken, May 2015. • S.-E. Walker: Model selection and focused scores a good tool for visualizing high-di- inference using robust estimators. Poster • M. Jullum: An approximate Bayesian mensional data? Oberwolfach workshop, & flashtalk, NordStat 2016, København, geophysical inversion framework based on February 2016. local-Gaussian likelihoods. PhD and Post- June 2016. • C. Cunen: When did Author Two take over Doc workshop at Klækken, May 2015. • C. Cunen: When did Author B take over for for Author One? Poster presentation, Author A? PhD Day, Oslo, June 2016. • C. Cunen: Optimal inference for two-by- Fusion Learning & BFF, Rutgers, April 2016. two tables modelled as Poisson pairs: • N.L. Hjort: Confidence distributions and • N.L. Hjort: Confidence distributions for a confidence distribution approach.18th objective Bayes. ISBA World Meeting 2016, change-points and regime shifts. Invited Norwegian Statistical Meeting, Solstrand, special topics session organised by Nils, talk, Fusion Learning & BFF, Rutgers, April June 2015. Sardegna, June 2016. 2016. • G.H. Hermansen: Predicting with con- • N.L. Hjort: Combining Information Across • N.L. Hjort: Confidence, Likelihood, Prob- fidence in time series processes. 18th Diverse Sources via Confidence Distribu- ability. Ten-hour short-course, University Norwegian Statistical Meeting, Solstrand, tions: the II-CC-FF Paradigm. Invited talk, Bicocca, Milan, April 2016. June 2015. JSM2016, Chicago, August 2016. • N.L. Hjort: The world’s first novel, skiing • N.L. Hjort: On just how surprised we ought • N.L. Hjort: Post-Selection Distributions, days near Oslo, Kola temperatures and the to have been when Bolt set his world Model Averaging, Bagging, Inference. Hjort time series. University Bocconi, Milan, records. 18th Norwegian Statistical Meet- Invited talk, post-selection inference work- April 2016. ing, Solstrand, June 2015. shop, Leuven, August 2016. • C. Cunen: Focused model selection for • M. Jullum: Parametric or nonparametric: • K.H. Hellton: Fridge: a focused approach meta-analysis of 2 x 2 tables. FICology The FIC approach for time series. 18th to fine-tuning ridge regression. PhD and workshop, Oslo, May 2016. Norwegian Statistical Meeting, Solstrand, PostDoc workshop at Klækken, September June 2015. • K.H. Hellton: The fridge: focused tuning of 2016. ridge regression. FICology workshop, Oslo, • C. Cunen: Optimal inference via confidence • E.Aa. Stoltenberg: Estimating proportions May 2016. distributions for two-by-two tables modelled from censored data. PhD and PostDoc as Poisson pairs [joint work with N.L. Hjort]. • G.H. Hermansen: Focused information workshop at Klækken, September 2016. criteria for time series (with N.L. Hjort). 60th ISI Conference, Rio, July 2015. • C. Cunen: When did author B take over for FICology workshop, Oslo, May 2016. • K.H. Hellton: Consistency of principal author A? Confidence Distributions and component scores in visualization of • N.L. Hjort: The hybrid likelihood: Combining Change-Points. PhD and PostDoc work- high-dimensional data [joint work with M. parametrics and nonparametrics. FICology shop at Klækken, September 2016. workshop, Oslo, May 2016. Thoresen]. 60th ISI Conference, Rio, July • K.H. Hellton: The Fridge. The L^\eta, 2015. • M. Jullum: FIC with a nonparametric candi- Многая лета FocuStat Research Kitchen, • G.H. Hermansen: Parametric or nonpara- date: A new strategy for FIC constructions. Oslo, October 2016. FICology workshop, Oslo, May 2016. metric: the FIC approach for stationary time • G.H. Hermansen: Focused Regularised series [joint work with M. Jullum and N.L. • S.-E. Walker: Focused inference based on Likelihood. The L^\eta, Многая лета Hjort]. 60th ISI Conference, Rio, July 2015. robust estimators (with N.L. Hjort). FICology FocuStat Research Kitchen, Oslo, October • N.L. Hjort: Combining information across workshop, Oslo, May 2016. 2016. diverse sources via confidence distributions. • C. Cunen: The world’s first novel: When did • N.L. Hjort: Minimum dispair, maximum 60th ISI Conference, Rio, July 2015. Author B take over for Author A? Poster & despair. The L^\eta, Многая лета FocuStat • S.-E. Walker: Model selection and verifica- flashtalk (winning one of three Best Poster Research Kitchen, Oslo, October 2016. Awards), NordStat 2016, København, June tion for ensemble based probabilistic fore- • S.-E. Walker: Minimum disparity estima- 2016. casting of air pollution in Oslo, Norway [joint tors. The L^\eta, Многая лета FocuStat work with G.H. Hermansen and N.L. Hjort]. • K.H. Hellton: Integrative clustering of Research Kitchen, Oslo, October 2016. 60th ISI Conference, Rio, July 2015. high-dimensional data with joint and indi-

41 TALKS & PRESENTATIONS

• C. Cunen: Statistikk for the Game of Despair. FocuStat BBB workshop, May Thrones. Fagligpedagogisk dag, University 2017. of Oslo, November 2016. • N.L. Hjort: Building Bridges (at Bislett): general workshop themes. FocuStat BBB workshop, May 2017. 2017: • V. Ko: Focused information criteria for cop- ulae. FocuStat BBB workshop, May 2017. • K.H. Hellton: Pop-Mat: 7 søyler med statis- • E.Aa. Stoltenberg: Inference in misspeci- tisk visdom. Matematisk fagutvalg, UiO, fied cure models. FocuStat BBB workshop, January 2017. May 2017. • N.L. Hjort: Model selection and model • S.-E. Walker: Focused model selection and averaging (2 hrs). Geilo Winter School, inference using robust estimators (with N.L. January 2017. Hjort). FocuStat BBB workshop, May 2017. • N.L. Hjort: Confidence distributions and • C. Cunen: Are the death rates in Game of data fusion (2 hrs). Geilo Winter School, Thrones realistic? A statistical case-study. January 2017. Oslo Data Science Meetup, May 2017. • N.L. Hjort: Bayesian nonparametrics (2 • C. Cunen: Combining diverse information Cunen gives her invited talk on combination hrs). Geilo Winter School, January 2017. sources with the II-CC- FF paradigm. The of diverse information sources, under the high patronage of His Majesty King Moham- • C. Cunen: The world’s first novel: When 19th national meeting of the NSF, Fredriks- med VI, at the 61st ISI World Congress in did Author B take over for Author A? Poster tad, June 2017. Marrakech, July 2017. presentation, Geilo Winter School, January • K.H. Hellton: Fridge: focused fine-tuning of 2017. ridge regression for personalized prediction. prepared (see the link), but not actually • V. Ko: Interview session about the book Met The 19th national meeting of the NSF, given; Cunen chose not to take part (see her hartelijke groente, with Wim Daniëls, Wim Fredrikstad, June 2017. statement of page 2); Hjort was ill on that Daniëls boekt, Boxmeer, January 2017. • G.H. Hermansen: Focused Regularised day of the conference.] • N.L. Hjort: A gentle introduction to Bayesian Likelihood. The 19th national meeting of the • S.-E. Walker (with N.L. Hjort): Estima- Nonparametrics. Big Insight Lunch Talk, NSF, Fredrikstad, June 2017. tion and model selection by data-driven February 2017. • N.L. Hjort: The semifinals factor for skiing weighted likelihood. Poster and speed fast in the finals. The 19th national meeting • K.H. Hellton: p-verdier. Metodemøte, presentation, JSM, Baltimore, July-August of the NSF, Fredrikstad, June 2017. Nasjonal kompetansetjeneste for kvinne- 2017. helse, February 2017. • E.Aa. Stoltenberg: Selecting the failure • C. Cunen: Game of Thrones vs the Wars of time distribution in cure models. The 19th • K.H. Hellton: p-values: what’s the problem? Roses: Survival and competing risks. Sylvia national meeting of the NSF, Fredrikstad, Symposium of the IMB PhD Forum, Feb- Richardson Honoris Causa workshop, UiO, June 2017. ruary 2017. August 2017. • S.-E. Walker: Estimation and model • C. Cunen: Er Game of Thrones blodigere • K.H. Hellton: Personalized prediction selection by data-driven maximum weighted enn middelalderkriger? An «åpen dag» through focused fine-tuning of ridge regres- likelihood. The 19th national meeting of the University of Oslo lecture, also shown (and sion. Sylvia Richardson Honoris Causa NSF, Fredrikstad, June 2017. kept) on television, NRK Skole, April 2017. workshop, UiO, August 2017. • N.L. Hjort: Empirical Likelihood and Bayes- • N.L. Hjort: Data fusion with confidence • N.L. Hjort: All models are wrong, but some ian Nonparametrics. Invited talk, the 11th curves: The II-CC-FF paradim. The BFF4 are more biologically plausible than others: Bayesian Nonparametrics Conference, conference, Harvard University, May 2017. Survival analysis models based on crossing Paris, June 2017. • C. Cunen: Analysis of fat weights: Three times for cumulative damage processes. new wide models. Presentation for the • G.H. Hermansen: Бернштейн–von Mises Sylvia Richardson Honoris Causa work- Scientific Committee of the International theorems for a class of Bayesian nonpara- shop, UiO, August 2017. metric setups for stationary time series. Whaling Commision, Bled, May 2017. • E.Aa. Stoltenberg: Some insight into the Poster presentation [with hours of wine and • C. Cunen: No paradox: Explaining how we cure model. Sylvia Richardson Honoris shouting], the 11th Bayesian Nonparamet- can observe a decrease in fat weight, while Causa workshop, UiO, August 2017. rics Conference, Paris, June 2017. the total weight appears to remain constant. • C. Cunen: Combining diverse informa- • C. Cunen: Combining diverse informa- Presentation for the Scientific Committee of tion sources with the II-CC-FF paradigm. tion sources with the II-CC-FF paradigm: the International Whaling Commision, Bled, Statistics seminar, MRC Biostatistics Unit, Independent Inspection, Confidence May 2017. Cambridge, September 2017. Conversion, Focused Fusion. Invited talk, • C. Cunen: Decline in energy storage in the 61st ISI World Conference, Marrakech, • C. Cunen: Død, vold og statistikk: Game of Antarctic Minke whales during the JARPA Marocco, July 2017. Thrones vs. Rosekrigene. Kaplah! Vol. 5, period: Assessment via the Focused Infor- Tilt, Oslo, October 2017. mation Criterion (FIC) [with Lars Walløe and • K.H. Hellton (with N.L. Hjort): Focused • V. Ko: Interview session about the book Met Nils Lid Hjort]. Presentation for the Scientific Fine-tuning of Ridge Regression for Person- hartelijke groente, Library Alphen aan den Committee of the International Whaling alized Prediction. Contributed paper and Rijn, October 2017. Commision, Bled, May 2017. presentation, JSM, Baltimore, July-August 2017. • N.L. Hjort: Whales, politics, and stat- • C. Cunen: Confidence distributions: A poten- isticians. Fagligpedagogisk dag, UiO, tial bridge for combining parametric and • N.L. Hjort (with C. Cunen and G. Herman- November 2017. nonparametric analyses. FocuStat BBB sen): Confidence distributions for change- workshop, May 2017. points and regime shifts. Invited talk, in an • E.Aa. Stoltenberg: Two approaches to Invited session on Complex risk structures statistics. Fagligpedagogisk dag, UiO, • K.H. Hellton: Focused Fine-tuning of Ridge or constraints of reliability of systems; JSM, November 2017. Regression for Personalised Prediction. Baltimore, July- August 2017. FocuStat BBB workshop, May 2017. • C. Cunen: Gamma process models for • C. Cunen (with N.L. Hjort): Combining competing risks. The From Processes to • G.H. Hermansen: Focused regularised like- diverse information sources with the Models FocuStat Research Kitchen, Oslo, lihood. FocuStat BBB workshop, May 2017. II-CC-FF paradigm: New developments. November 2017. • N.L. Hjort: The hybrid likelihood. FocuStat Invited talk, in an Invited session, organised • N.L. Hjort: From Processes to Models: a BBB workshop, May 2017. by Nils, on BFF in Action; JSM, Baltimore, mini-review. The From Processes to Models • N.L. Hjort: Minimum Dispair, Maximum July-August 2017. [The talk was well

42 TALKS & PRESENTATIONS

FocuStat Research Kitchen, Oslo, Novem- that’s the question [joint work with Nils ber 2017. Lid Hjort]. FocuStat VVV Conference, • N.L. Hjort: Some Gamma process based Ingeniørenes Hus, May 2018. models for survival data. The From • V. Ko: FIC for copulae [joint work with Nils Processes to Models FocuStat Research Lid Hjort]. FocuStat VVV Conference, Kitchen, Oslo, November 2017. Ingeniørenes Hus, May 2018. • R. Parviero: Playing basketball via simu- • T. Schweder: Unbiased confidence. lated Markov chains. The From Processes FocuStat VVV Conference, Ingeniørenes to Models FocuStat Research Kitchen, Hus, May 2018. Oslo, November 2017. • E.Aa. Stoltenberg: Models and inference • E.Aa. Stoltenberg: Cure models and for on-off data via clipped Ornstein-Uhlen- path dependent hazard rates. The From beck processes [joint work with N.L. Hjort]. Processes to Models FocuStat Research FocuStat VVV Conference, Ingeniørenes A biased subset of Statisticians Against War, Kitchen, Oslo, November 2017. Hus, May 2018. at BI, November 2018: Hermansen, Haug, • C. Cunen: Combining diverse information • C. Cunen: II-CC-FF: Combination of infor- Cunen, Hjort, with Cunen and Hjort giving sources with the II-CC-FF paradigm, with mation beyond meta-analysis [joint with with an invited joint seminar talk. applications in meta-analysis & beyond. N.L. Hjort]. NordStat, Tartu (with Cunen as London School of Hygiene & Tropical invited organiser for a session on CDs and • N.L. Hjort (with F. Aakre): Dolce and Medicine, November 2017. related themes), June 2018. Largo, from Telemann’s F major duet for alto recorders. FocuStat Combo Kitchen, • K.H. Hellton: Significance for a modern • G.H. Hermansen: Peacekeeping operations November 2018. world. Big Insight Lunch Talk, Norwegian and the intensity of violence in internal Computing Centre, December 2017. armed conflict [joint work with H. Nygård]. • C. Cunen: II-CC-FF – combination of NordStat, Tartu, June 2018. information beyond meta-analysis. FocuStat Combo Kitchen, November 2018. • N.L. Hjort: Prediction and confidence. NordStat, Tartu, June 2018. • K.H. Hellton: Combination of information 2018: across diverse sources: An overview. • V. Ko: Focused information criterion for FocuStat Combo Kitchen, November 2018. copula [joint work with N.L. Hjort]. NordStat, • N.L. Hjort: Discussion of Clauset’s `Trends Tartu, June 2018. • C. Cunen: Freedman’s Paradox. Trial lec- ture for her PhD disputation. Department of and Fluctuations’. PRIO, Oslo Lectures on • E.Aa. Stoltenberg: Models and inference Mathematics, University of Oslo, December, Peace and Conflict, January 2018. for on-off data via clipped Ornstein-Uhlen- ­2018. • G.H. Hermansen: Big data: Hva er likheten beck processes [joint work with N.L. Hjort]. mellom epigenetikk og bildegjenkjenning? NordStat, Tartu, June 2018. • C. Cunen: Wars and Whales: Extensions and Applications of Confidence Curves and The University of Oslo’s Åpen Dag, March • S.-E. Walker: Focused model selection and Focused Model Selection. Introduction to 2018. inference using robust estimators [joint work the public attack-and-defense disputation of • N.L. Hjort: One Thousand is Unfair, Two with N.L. Hjort]. NordStat, Tartu, June 2018. her PhD dissertation. Department of Math- Thousand is Fair. Matematisk felleskol- • K.H. Hellton: Fridge: Focused fine-tun- lokvium, Department of Mathematics, UiO, ematics, University of Oslo, December, ing of ridge regression for personalized 2018. March 2018. predictions [joint work with Nils Lid Hjort]. • K.H. Hellton: The Fridge: focused fine-tun- • F. Ferraccioli: Nonparametric penalized CompStat, Iași, Romania, August 2018. ing of ridge regression for personalised pre- likelihood for density estimation. Statistics • C. Cunen: Når startet Den lange freden? seminar talk, Department of Mathematics, dictions. Computational and Methodological [joint work with N.L. Hjort and H. Nygård] Statistics, Pisa, December, 2018. UiO, April 2018. Invited opening talk at the Annual Alumni • E.Aa. Stoltenberg (with N.L. Hjort): Models Celebrations of the Department of Mathe- and inference for on-and-off data via clipped matics, University of Oslo, September 2018. Ornstein-Uhlenbeck processes. Poster • C. Cunen: Game of Thrones and Reality. presentation, New Aspects of Statistics, PRIO, one of three lectures, followed by Financial Econometrics, and Data Science. a one-hour panel discussion, also with Spring Conference, Stevanovich Center, questions from the plus-hundred audience. University of Chicago, May 2018. September 2018. • C. Cunen: Whales, politics, and statisticians • C. Cunen, with N.L. Hjort and H. Nygård: [joint work with N.L. Hjort and L. Walløe]. Statistical Sightings of Better Angels. FocuStat VVV Conference, Ingeniørenes Brown-bag G-35 lunch seminar at PRIO, Hus, May 2018. discussing the paper. October 2018. • K.H. Hellton: High-dimensional asymptotics • C. Cunen: II-CC-FF – Combination of of principal component regression. FocuStat information beyond meta-analysis [joint VVV Conference, Ingeniørenes Hus, May work with N.L. Hjort]. PhD and PostDoc 2018. workshop at Klækken, October 2018. Brad Efron, Regina Liu, and N.L. Hjort, • G.H. Hermansen: Bayesian nonparametrics • N.L. Hjort: War, Peace, and Statisticians. invited speakers at the BFF. for stationary time series [joint work with Fagligpedagogisk dag, UiO, November N.L. Hjort]. FocuStat VVV Conference, 2018. Ingeniørenes Hus, May 2018. • C. Cunen and N.L. Hjort: Statistical Sight- • N.L. Hjort: Welcoming remarks & introduc- ings of Better Angels. Statistics seminar (in tion to the FocuStat VVV themes (touching stereo), BI Norwegian Business School, also Lessons A, B, C, D from the Birke- November 2018. land-Eyde world). FocuStat VVV Confer- ence, Ingeniørenes Hus, May 2018. • K.H. Hellton: Focused fine-tuning of ridge regression with data integration. FocuStat • N.L. Hjort: Survival and event history Combo Kitchen, November 2018. models via Gamma processes [joint work with C. Cunen]. FocuStat VVV Conference, • N.L. Hjort: Olympic unfairness – analysing Ingeniørenes Hus, May 2018. 2 x 1000 m speedskating races. FocuStat Combo Kitchen, November 2018. • M. Jullum: Parametric or nonparametric,

43 PUBLICATIONS

Publications Books, journal articles, technical reports, blog posts, pamphlets, Publications by the FocuStat fliers, FocuStat Notes, libretti, participants, including also those statistical poems, symposiae from 2013 of relevance proceedings, and yet further names of the publication games: Here we for the 2014-2018 project. list publications by the FocuStat participants, including also those from 2013 of relevance for the 2014-2018 project.

• Argyrou, J. (2017). Topics in Confidence Introduction chapter (the so-called kappen) Joint Statistical Meeting 2016, the American Distributions. Master’s thesis (supervisor: to her PhD dissertation, Department of Statistical Association, 138–153. The JSM N.L. Hjort), Department of Mathematics, Mathematics, University of Oslo, Decem- took place in Chicago, July 29 to August 4, University of Oslo. ber 2018. where Hjort was an invited speaker in the • Barua, S. (2017). Inverse covariance matrix • Cunen, C. (2018). Wars and Whales: specially invited session ‘Bridging Bayes- estimation for the global minimum variance Extensions and Applicatiaons of Confidence ian/Frequentist/Fiducial Inference in the portfolio. Master’s Thesis, Department of Curves and Focused Model Selection. The Era of Data Science’. Mathematics, UiO (supervisors: I.K. Glad full PhD dissertation (supervisor: N.L. Hjort), • Cunen, C., Hjort, N.L. (2018). Combining and N.L. Hjort). defended December 2018, Department of information across diverse sources: the • Bjørvik, L.M., Dale, S., Hermansen, G.H., Mathematics, University of Oslo. II-CC-FF paradigm. Statistical Research Munishi, P.K.T., Moe, S.R. (2015). Bird flight • Cunen, C., Hermansen, G.H., Stoltenberg, Report, Department of Mathematics, UiO, initiation distances in relation to distance E.Aa. (2017). Bokmelding: T. Schweder and September 2018 (submitted for publication). from human settlements in a Tanzanian N.L. Hjort’s Confidence, Likelihood, Proba- • Cunen, C., Hjort, N.L., Nygård, H. (2018). floodplain. Journal of Ornithology, 156, bility (Cambridge University Press, 2016). Statistical Sightings of Better Angels: 239-246. FocuStat Blog Post, November 2017. Analysing the Distribution of Battle Deaths • Cunen, C. (2014). Three PLS-based • Cunen, C., Hermansen, G., Stoltenberg, in Interstate Conflict over Time. Journal of methods for variable ranking and interaction E.Aa. (2017). Book review: T. Schweder Peace Research [to appear]. detection. Master thesis, Norwegian Univer- and N.L. Hjort’s Confidence, Likelihood, • Cunen, C., Walløe, L., Konishi, K., Hjort, sity of Life Sciences. Probability: Statistical Inference With N.L. (2018). Supplementary notes and • Cunen, C. (2015). Mortality and Nobility Confidence Distributions (Cambridge material, with some refined analyses, com- in the Wars of the Roses and Game of University Press, 2016). Tilfeldig Gang, nr pared to our IWC/SC/67A/EM04 May 2017 Thrones. FocuStat Blog Post, October 1, juni 2017, pages 9-12. report. Report to the Scientific Committee 2015. • Cunen, C., Hermansen, G.H., Hjort, N.L. of the International Whaling Commission, Bled, May 2018: SC/67/EM/02. • Cunen, C. (2015). The impact of philosophy (2018). Confidence distributions for change- of science: A citation analysis. Exam essay, points and regime shifts. Journal of Statisti- • Cunen, C., Walløe, L., Hjort, N.L. (2017). Science, Ethics, Society PhD course mnses cal Planning and Inference 195, 14-34. Decline in energy storage in Antarctic Minke 9100, December 2015. • Cunen, C., Hjort, N.L. (2015). Optimal whales during the JARPA period: Assess- inference via confidence distributions for ment via the Focused Information Criterion • Cunen, C. (2017). Analysis of fat weights: (FIC). Reports of the Scientific Committee Three new wide models. Reports of the two-by-two tables modelled as Poisson pairs: Fixed and random effects. [For the of the International Whaling Commision Scientific Committee of the International SC/67A/EM/04. Whaling Commision SC/67A/EM/08. Rio 60th World Statistics Congress.] • Cunen, C., Hjort, N.L. (2017). New statistical • Cunen, C., Walløe, L., Hjort, N.L. (2018). • Cunen, C. (2017). No paradox: Explaining Reactions and answers to two papers by how we can observe a decrease in fat methods shed light on medieval literary mystery. FocuStat Blog Post, January 2017. McKinlay, de la Mare and Welsh. Report to weight, while the total weight appears to the Scientific Committee of the Interna- remain constant. Reports of the Scientific • Cunen, C., Hjort, N.L. (2018). Whales, tional Whaling Commission, Bled, May Committee of the International Whaling Politics, and Statisticians. FocuStat Blog 2018: SC/67B/EM/08. Commision SC/67A/EM/07. Post, January 2018. • Cunen, C., Walløe, L., Hjort, N.L. (2018). • Cunen, C. (2018). Wars and Whales: • Cunen, C., Hjort, N.L. (2016). Combining Focused model selection for linear mixed Extensions and Applications of Confidence information across diverse sources: The models, with an application to whale ecol- Curves and Focused Model Selection. II-CC-FF paradigm. Proceedings from the ogy. Submitted for publication.

44 PUBLICATIONS

• Garcia-Magarios, M., Egeland, T., and why are principal component scores a • Hjort, N.L. (2017). Cooling of Newborns and Lopez-de-Ullibarri, Hjort, N.L., Salas, A. good tool for visualizing high-dimensional the Difference Between 0.244 and 0.278. (2015). A parametric approach to kinship data? Scandinavian Journal of Statistics, FocuStat Blog Post, December 2017. hypothesis testing using identity-by-de- 44, 581-597. • Hjort, N.L. (2018). Towards a More Peace- scent parameters. Statistical Applications • Hellton, K.H. (2015). The bonus roll and ful World [insert ‘!’ or ‘?’ here]. FocuStat in Genetics and Molecular Biology, 14, the Christmas spirit. FocuStat Blog Post, Blog Post, January 2018. 465-479. December 2015. • Hjort, N.L. (2018). One Thousand is • Giil, L.M., Solvang, S.-E. H., Giil, M.M., • Hermansen, G.H. (2014). Model Selection Unfair, Two Thousand is Fair. Pre-Olympic Hellton, K.H., Skogseth, R.E., Vik-Mo, and Bayesian Nonparametrics for Time FocuStat Blog Post, February 2018. Hjort, A.O., Hortobagyi, T., Aarsland, D., Nor- Series and Nonstandard Regression Mod- N.L. (2018). The Best Metal-Grabbing drehaug, J.E. (2018). Serum potassium els. PhD dissertation (supervisor: N.L. Hjort), Games Ever. Post-Olympic FocuStat Blog is associated with cognitive decline in defended September 2014, Department of Post, February 2018. patients with Lewy Body Dementia. Journal Mathematics, University of Oslo. of Alzheimer’s Disease. • Hjort, N.L. (2018). Belgium Breaks a 48 • Hermansen, G.H., Hjort, N.L. (2015). Bern- Year Old Curse (Apparently). FocuStat • Giil, L., Aarsland, D., Hellton, K.H., Lund, stein-von Mises theorems for nonparametric Blog Post, July 2018. Hjort, N.L. (2018). A., Heidecke, H., Schulze-Forster, K., function analysis via locally constant model- Overdispersed Children. FocuStat Blog Riemekasten, G., Vik-Mo, A.O., Kristof- ling: A unified approach. Journal of Statistical Post, August 2018. fersen, E.K., Vedeler, C.A., Nordrehaug, Planning and Inference, 166, 138-157. J.E. (2018). Antibodies to multiple recep- • Hjort, N.L. (2016). Exercises & Lecture tors are associated with neuropsychiatric • Hermansen, G.H., Hjort, N.L. (2014). Notes STK 4180, for the master-and-PhD symptoms and mortality in Alzheimer’s Focused information criteria for time series. course on Confidence Distributions. disease: a longitudinal study. Journal of Statistical Research Report, Department of • Hjort, N.L. (2017). Exercises & Lecture Alzheimer’s Disease, 64, 761-774. Mathematics, University of Oslo. Notes STK 4021, for the master-and-PhD • Glad, I.K., Hjort, N.L. (2016). Model uncer- • Hermansen, G.H., Hjort, N.L., Jullum, M. course on Model Selection and Model tainty first, not afterwards. Discussion con- (2015). Parametric or nonparametric: The Averaging. tribution to the article ‘Approximate models FIC approach for stationary time series. [For • Hjort, N.L. (2017). Exercises & Lecture and robust decisions’ by J. Watson and C. the Rio 60th World Statistics Congress.] Notes STK 4160, for the master-and-PhD Holmes. Statistical Science 31, 490-494. • Hermansen, G., Hjort, N.L., Jullum, M. course on Model Selection and Model • Grønneberg, S., Hjort, N.L. (2014). The (2015). Parametric or nonparametric: The Averaging. Copula Information Criteria. Scandinavian FIC approach for time series. Technical • Hjort, N.L. (2018). Exercises & Lecture Journal of Statistics, 41, 436-459. report, Department of Mathematics, Notes STK 9190, for the master-and-PhD University of Oslo. [Extended and further course on Bayesian Nonparametrics. • Hauge, V.L., Hermansen, G.H. (2017). developed 21-page version of a six-page Machine Learning methods for sweet spot story we’ve published in the ISI 2015 Rio • Hjort, N.L. (2018). Exercises & Lecture detection: a case study. Quantitative Geol- de Janeiro proceedings.] Notes STK 4080, for the master-and-PhD ogy and Geostatistics, 19, 573-588. course on Survival Analysis and Event • Hermansen, G., Hjort, N.L., Kjesbu, O. History Analysis. • Hellton, K.H. (2014). On High-Dimensional (2016). Recent advances in statistical Principal Component Analysis in Genomics: methodology applied to the Hjort liver index • Hjort, N.L., McKeague, I.W., Van Keilegom, Consistency and Robustness. PhD Disser- time series (1859-2012) and associated I. (2018). Hybrid combinations of parametric tation (supervisors: M. Thoresen and A. influential factors. Canadian Journal and empirical likelihoods. Statistica Sinica, Frigessi), Department of Biostatistics, UiO. of Fisheries and Aquatic Sciences 73, 28, 2389-2407. Complete with a Supple- • Hellton, K.H. (2015). Hvor stemmes det 279–295. This is part of the special issue mentary Material online section. Our paper hva i hovedstaden? FocuStat Blog Post, based on the Johan Hjort Symposium on is for Sinica’s Peter Hall volume, and in September 2015. Hellton, K.H. (2015). The Recruitment Dynamics and Stock Variabil- the introduction section Hjort also reports bonus roll and the Christmas spirit. FocuStat ity, Bergen, Norway, 7–9 October 2014. briefly on Peter’s visit to Oslo in 2003. Blog Post, December 2015. • Hjort, N.L. (2014). Discussion of Efron’s • Hjort, N.L., Schweder, T. (2018). Confidence • Hellton, K.H., Cummings, J., Vik-Moe, A.O., ‘Estimation and accuracy after model selec- distributions and related themes. General Nordrehaug, J.E., Aarsland, D., Selbaek, G., tion’. Journal of the American Statistical introduction article to a Special Issue of the Giil, L.M. (2018). The truth behind the zeros: Association, 110, 1017-1020. Journal of Statistical Planning and Inference A new approach to Principal Component dedicated to this topic, with eleven articles, • Hjort, N.L. (2014). Deilig er jorden. In Ad and with Hjort and Schweder as guest Analysis of the neuropsychiatric inventory. Fontes, festschrift for Terje Kvam, 101-103. Submitted for publication. editors; vol. 195, 1-13. • Hjort, N.L. (2015). Norway-Spain 28-25: • Hellton, K.H., Hjort, N.L. (2018). Fridge: • Hjort, N.L., Sørensen, Ø. (2018). Rikdom og How exciting was it? FocuStat Blog Post, pengenes verdi i Andeby (Being rich, and Focused Finetuning of Ridge Regression September 2015. for Personalized Predictions. Statistics in the value of money, in Duckburg). Donald Medicine 37, 1290-1303. • Hjort, N.L. (2015). Conspiracy Probabil- Duck, De komplette årgangene (volume ity Calculus for Norwegian Footballers. 1976, part VI), Egmont, Walt Disney, • Hellton, K.H., Røislien, J. (2017). Verdens FocuStat Blog Post, October 2015. November 2018. første p-verdi. Tidsskrift for Den norske legeforening, årgang 137, mai 2017. • Hjort, N.L. (2016). Recruitment Dynamics • Hjort, N.L., Walker, S.-E. (2019). Estima- and Stock Variability: The Johan Hjort tion and model selection via weighted • Hellton, K.H., Thoresen, M. (2014). The Symposium, some personal reflections. likelihoods. Manuscript. impact of measurement error on principal FocuStat Blog Post, April 2016. component analysis. Scandinavian Journal • Jullum, M. (2016). New Focused of Statistics, 41, 1051-1063. • Hjort, N.L. (2016). Rapport fra NordStat Approaches to Topics Within Model Selection 2016, København 27.-30. juni. In Tilfeldig and Approximate Bayesian Inverstion. PhD • Hellton, K.H., Thoresen, M. (2015). Con- Gang, December 2016 issue, pages 7-8. dissertation (supervisors: N.L. Hjort and O. sistency of principal component scores in • Hjort, N.L. (2017). But some are more Kolbjørnsen), defended April 2016, Depart- visualizations of high-dimensional data. [For ment of Mathematics, University of Oslo. the Rio 60th World Statistics Congress.] equal than others. FocuStat Blog post, January 2017. • Jullum, M., Baños, D.R., Hjort, N.L. (2015). • Hellton, K.H., Thoresen, M. (2016). Integra- To liv: kvinnene i Lillestrøm som ble født tive clustering of high-dimensional data with • Hjort, N.L. (2017). The semifinals factor for skiing fast in the finals. FocuStat Blog Post, på samme dag og døde på samme dag. joint and individual clusters. Biostatistics FocuStat Blog Post, October 2015. 17, 537-548. February 2017. Hjort, N.L. (2017). Ut på tur: NordStat, København, juni 2016. FocuStat • Jullum, M., Hjort, N.L. (2017). Parametric or • Hellton, K.H., Thoresen, M. (2017). When Blog Post, February 2017. nonparametric: The FIC approach. Statis-

45 PUBLICATIONS

tica Sinica, 27, 951–981. With Supplemen- ing the effect of meteorology on trends in (2018). Bayesian regional flood frequency tary Material (19 pages). surface ozone: Development of statistical analysis for large catchments. Water • Jullum, M., Hjort, N.L. (2018). What price tools. ETC/ACM Technical Paper. Resource Research. semiparametric Cox regression? Lifetime • Stoltenberg, E. Aa. (2015). The c-Loss • Vanem, E., Walker, S.E. (2013). Identifying Data Analysis, 24, 1-33. With Supplemen- Function: Balancing Total and Individual trends in the ocean wave climate by time tary Material (7 pages). Risk in the Simultaneous Estimation of series analyses of significant wave height • Jullum, M., Kolbjørnsen, O. (2016). A Poisson Means. Master Thesis in statistics data. Ocean Engineering, 61, 148-160. Gaussian based framework for local Bayes- (supervisor: N.L. Hjort), Department of • Vonnet, J., Hermansen, G.H. (2015). Using ian inversion of geophysical data to rock Mathematics, University of Oslo. predictive analytics to unlock unconven- properties. Geophysics, 81, 75-87. • Stoltenberg, E.Aa. (2017). Objectivity, tional plays. First Break, 33, 87-92. • Ko, V., Wieling, M., Wit, E., Nerbonne, J., rationality and statistics. Exam essay for the • Walker, S.-E. (2015). Noen etiske problem- Krijnen, W. (2014). Social, geographical, Science, Ethics, Society PhD course mnses stillinger i forbindelse med håndtering og and lexical influences on Dutch dialect 9100, December 2017. formidling av usikkerhet i miljøforskning. pronunciations. Computational Linguistics • Stoltenberg, E. Aa. (2017). Frekventisme, Exam essay, Science, Ethics, Society PhD in the Netherlands Journal, 4, 29-38. Bayes og likelihoodprinsippet. Filosfisk sup- course mnses 9100, December 2015. • Ko, V. (2016). Met hartelijke groente. plement, 2 (special issue on the Philosophy • Walker, S.-E., Denby, B. R., Castell, N., Nederlands door Koreanse ogen. Uitgeverij of Mathematics), 58-65. Pross, B., Cornford, D. (2014). Ensemble Podium, Amsterdam (175 pages). [No, • Stoltenberg, E.Aa. (2017). Frekventisme, based probabilistic forecasting of air quality FocuStat can’t take any credit for this at all, Bayes og likelihoodprinsippet. FocuStat in Oslo and Rotterdam. Technical Report, but we feel Vinnie’s book belongs on our Blog Post, July 2017. NILU, Norway. publication list, nevertheless.] • Stoltenberg, E.Aa., Hjort, N.L. (2015). The • Walker, S.-E., Hermansen, G.H., Hjort, N.L. • Ko, V., Hjort, N.L. (2019). Model robust infer- c-loss function: Balancing total and indi- (2015). Model selection and verification for ence with two-stage maximum likelihood vidual risks in the simultaneous estimation ensemble based probabilistic forecasting estimation for copulae. Journal of Multivari- of Poisson means. Proceedings of the 19th of air pollution in Oslo, Norway. [For the Rio ate Analysis [to appear]. European Young Statisticians Meeting, 60th World Statistics Congress.] • Ko, V., Hjort, N.L. (2019). Copula information September 2015, Prague, 133-138. • Walker, S.-E., Hjort, N.L. (2019). Focused criterion for model selection with two-stage • Stoltenberg, E.Aa., Hjort, N.L. (2018). Mod- model selection and inference using robust maximum likelihood estimators. Economet- els and inference for on-off data via clipped estimators. Statistical Research Report, rics and Statistics [to appear]. Ornstein-Uhlenbeck processes. Submitted Department of Mathematics, University of • Ko, V., Hjort, N.L., Hobæk Haff, I. (2019). for publication, October 2018. Oslo. Focused information criteria for copulae. • Stoltenberg, E.Aa., Nordeng, H.M.E., • Walløe, L., Hjort, N.L., Thoresen, M. (2018). Scandinavian Journal of Statistics [to Ystrøm, E., Samuelsen, S.O. (2018). Cure Major concerns about late hypothermia appear]. models and long term adverse outcomes study. Acta Paediatrica, November 2018. • Løland, A., Huseby, R.B., Hjort, N.L., after gestational exposure to pharmaceuti- • Østbye, P. (2014). Cost information-value Frigessi, A. (2013). Statistical corrections of cals. Submitted for publication, December trade-off in covariate selection. Master invalid correlation matrices. Scandinavian 2018. Thesis in statistics (supervisor: N.L. Hjort), Journal of Statistics, 40, 807-824. • Sørensen, Ø., Hellton, K.H., Frigessi, A., Department of Mathematics, University of • Moss, J. (2015). Cube Root Asymptotics. Thoresen, M. (2018). Covariate selection Oslo. Master Thesis in statistics (supervisor: in high-dimensional generalized linear • Øygard, A.M. (2015). Measuring similarity of N.L. Hjort), Department of Mathematics, models with measurement errors. Journal classified advertisements using images and University of Oslo. of Computational and Graphical Statistics text. Master Thesis in statistics (supervi- 27, 739-749. • Olsen, H.G., Hermansen, G.H.(2017). sors: G.H. Hermansen and N.L. Hjort), Recent advancements to nonparametric • Thorarinsdottir, T.L., Hellton, K.H., Stein- Department of Mathematics, University of nodeling of interactions between reservoir bakk, G.H., Engeland, K., Schlichting, L. Oslo. parameters. Quantitative Geology and Geostatistics, 19, 653-669. • Parviero, R. (2018). Improving Ridge Regression via Model Selection and Focussed Fine-Tuning. Tesi de laurea magistrale, Università degli Studi di Mila- no-Bicocca (supervisors: K.H. Hellton and N.L. Hjort), March 2018. • Rønneberg, L.T.S. (2017). Fiducial and Objective Inference: History, Theory, and Comparisons. Master thesis in statistics (supervisor: N.L. Hjort), Department of Mathematics, University of Oslo. • Schweder, T. (2017). Bayesian Analysis: Always and Everywhere? FocuStat Blog Post, November 2017. • Schweder, T., Hjort, N.L. (2013). Discussion of ‘Confidence distribution, the frequentist distribution estimator of a parameter: A review’. International Statistical Review, 81, 56-68. • Schweder, T., Hjort, N.L. (2016). Confidence, Likelihood, Probability. Statistical Inference with Confidence Distributions. Cambridge University Press (500 pages). • Solberg, S., Walker, S.-E., Schneider, P., Three PhD dissertations, with several more to come in 2019 and 2020, and the Special Issue Guerreiro, C., Colette, A. (2018). Discount- of Journal of Statistical Planning and Inference which started with our FocuStat Workshop “Inference With Confidence”.

46 Stoltenberg, Ko, Hjort, Cunen, Hermansen, “nils hjort with his latest book”, tweets Hjort, Cunen, Moss, Stoltenberg, Hellton, Parviero, this time at Litteraturhuset. Cambridge University Press, at JSM 2016. after Brahms A minor.

“with absolutely fantastic talks from Nils Hjort and Art Owen”, blogs Christian Nils blogs on Overdispersed and Markov- ­Robert from Paris. ian Children. TTT at the VVV.

Cunen, Hellton, Hermansen, still smiling, Another end-of-semester party. close to midnight, after a long working day. Inderdaad.

FocuStat wins the department quiz: Hell- ton, Ko, Cunen, Moss, Stoltenberg, Hobæk Walker finesses his plan for completing the Ready for twelve verses of FocuStats røst, Haff, Hjort. FRIC paper with Hjort. in D minor.

47 FocuStat after FocuStat

The Vårens Vakreste Variabler participants, May 2018.

WE HAVE BEEN privileged with continue to exist, as will in partic- ProcMod such data?, becomes an the generous funding awarded ular the FocuStat Blog pages, with operative research question. N.L. us by the Norwegian Research new Statistical Stories. The core Hjort shall be applying for funding Council. It has allowed us not group shall continue to meet reg- for a well-built ProcMod project, only woman- and manpower, qua ularly, with «runde rundt bordet» in that case hosting new PostDocs PostDocs and PhDs and associated and the necessary time to discuss and room for extensive collabora- colleagues and Master level stu- new ideas, partly using our favour- tion. Second, together with Håvard dents, but the flexibility to investi- ite technology of chalk, blackboard, Nygård at the Peace Research Insti- gate broader horizons and to build and sponge. tute Oslo, we’re seeking funding for lasting friendships and networks, Stability and Change. This project nationally and internationally. TWO IDENTIFIABLE LARGER aims at developing methodology In particular, our workshops and research themes on our desks and for understanding the complex pro- the VVV conference, along with blackboards are as follows. First, cesses of armed conflicts, as these our research kitchens, have been From Processes to Models, having develop over time, in country after fruitful, inspiring, and productive, to do with constructing classes country, influenced by a multitude leading to new ideas and to collabo- of plausible stochastic processes of interacting factors. What causes ration projects. “behind the data”, e.g. in biologi- change points and regime shifts, cal or biostatistical settings. Thus, what might be enough to secure IN SOME PRACTICAL, work- instead of applying off-the-shelf relative stability? The envisaged wise, and sociological senses the methods for a survival regression methodology will be eminently use- FocuStat project will live on, also dataset, for example, we delve into ful also for other application areas, as it will pave the way towards new the underlying mechanisms, to including biomedicine. themes, challenges, and constel- build better and more context-rel- lations. The FocuStat website will evant models for such data. Can we