ERC-Hjort Focus Driven Statistical Inference With Complex Data

Our project has lasted from Jan 2014 to Dec 2018. It has had PhDs Céline Cunen (from Aug 2014 to disputation Dec 2018) and Sam-Erik Walker (from Oct 2014 to Mar 2018; he completes his PhD in 2019), and PostDocs Gudmund Hermansen (from Oct 2014 to Mar 2017) and Kristoffer Hellton (from May 2015 to Dec 2018). They have all been significantly vital for the high quality, broad horizons, and carrying through of the project. In addition we have had other PhDs, supervised or co-supervised by Hjort, sufficiently close to the project that they have taken part in regular meetings, given talks at our workshops and research kitchens, etc.: Martin Jullum (PhD 2016), Vinnie Ko and Emil Stoltenberg (completing their dissertations in 2019; Jonas Moss and Martin Tveten (complete their in 2020). We've also had Italian half-a-year visitors (one master, one PhD). Finally, several of our colleagues, from Norway and abroad, have had close working relationships with some of us, for part of the project's duration. As such the project has had considerably larger volume and impact than "only" to support 2 PhDs plus 2 PostDocs. Our themes have been methodological, in several specified directions, but with a clear applied profile. For some of our work problems from substantive applied areas have been clear inspiration for methodological developments. A brief description of themes is as follows. 1. Confidence distributions, yielding inference tools for posterior distributions for the most vital parameters, without the Bayesian setup with priors (Cunen, Hjort, Hermansen, with Tore Schweder, Department of Economics). 2. Focused model building and model selection, partly via FIC, the focused information criterion (Cunen, Hellton, Hermansen, Hjort, Ko, Walker, with Lars Walløe). 3. Combination of information across diverse sources (Cunen, Hjort). 4. Robustification of certain classical procedures, for estimation and model selection (Hjort, Walker). 5. Change points, or regime-shifts (Cunen, Hermansen, Hjort, Tveten, with Håvard Nygård, Prio). This concerns reaching precise statements about whether a phenomenon has been stable over time, or whether there have indeed been change points. Methods aim at identifying such change points and also assessing the degree of change, with assigned levels of uncertainty. 6. Personalised inference, e.g. for biomedical applications (Hellton, Hjort). 7. Bayesian nonparametrics, with new types of applications (Hermansen, Hjort, Stoltenberg). Our applications have spanned broad horizons. Some of these have involved whales (Cunen, Hjort, with Walløe); wars and armed conflicts (Cunen, Hermansen, Hjort, Walker, with Nygård, and master student J.K. Haug); complex survival and event history analysis (Cunen, Hjort, Stoltenberg); literary analysis (Cunen, Hjort); macro economics (Hjort, with Schweder); the declining ozone levels over Europe (Hjort, Walker); sports (Hjort). We have a very active blog, with some 25 Statistical Stories, which shall continue to be used after Jan 2019. Some of these stories have reached far and wide, with thousands of readers (Steven Pinker's tweet about Hjort's "sophisticated analysis of battle death distributions" led to a thousand new, inside a week). This attention has also led to interviews in various media. Cunen won the Titan Prize 2018 for science communication, and has given a very high number of talks, presentations, and interviews. Of vital importance for us have been the three three-day international workshops, Inference With Confidence (May 2015), FICology, on focused model building and selection (May 2016), Building Bridges (May 2017), about bridging parametrics and nonparametrics. In May 2018 we organised a bigger four-day international summing-up conference, with 45 participants. The conferences have had a high international standard. Papers flowing from one of the workshops led to a Special Issue of the JSPI journal, with Hjort and Schweder as Guest Editors. We've also organised five so-called Research Kitchens, with international guests; in Sept 2014 on empirical likelihood, in 2015 on robust estimation via scoring rules and minimum divergence estimators; in Oct 2016 on robustified likelihoods, for different purposes; in Nov 2017 on From Processes to Models, and finally on Combination of Information in Nov 2018. Each year we've singled out one conference where we all went, to contribute actively, with talks etc. In 2015 we went to the ISI in Rio; in 2016 to the NordStat in København (with Cunen and Hellton winning best poster awards); in 2017 to the Norwegian meeting in Fredrikstad; and in 2018 again to NordStat, in Tartu. We have organised invited sessions etc. For an overview of published work, our 138 talks and presentations, our awards & prizes, occasions in the media, and blog stories, consult our web pages, which will continue to be active also after Jan 2019. We also recommend our separately published Summary Report (48 pages).

Her var jeg ikke klar over den strikte 1000-tegn-grensen, så min lange tekst på 4000 tegn får altså ikke plass. Her kommer derfor en meget amputert versjon. NFR kan få min fulle tekst om det er interesse for det. 1. Kompetanseheving, ikke bare for kjernegruppen, men for mange av de andre PhD- er og PostDoc-er ved statistikk UiO. 2. Privilegert og krevende samarbeid med eksperter innen havøkologi (Walløe) og krig-og-konfliktforskning (Nygård, Prio). Vi har vedvarende prosjekter med disse. 3. Et høyt antall publikasjoner, 138 foredrag og presentasjoner, mange populærvitenskapelige. Ved siden av mange Statistical Stories, publisert på vår blogg, og som har nådd tusener, har vi vært med på å sette fagstatistikk på kartet. 4. Hjort vil søke om videre prosjektstøtte, (a) om From Processes to Models, via FRIPRO, i april 2019; (b) om det tverrfaglige Stability and Change, med Nygård (PRIO).

Statistics is the science of reaching decisions under uncertainty and is in many respects a far-ranging success story, permeating nearly all substantive sciences and areas of society where data are collected. It has used around hundred years to reach its present state of high maturity and uniform usefulness. In broad strokes, the four main areas associated with (i)parametrics (models indexed by low-dimensional parameters); (ii) nonparametrics (models with high- or infinite-dimensional parameters); (iii) assessment and selection of models; and (iv) combination of different sources of information, drive most of modern statistics, and have, in essence, been well sorted out, conceptually and operationally. There are important gaps to be filled and new para digms and principles to develop, however, when faced with the statistical challenges of the 21st century. New types of data, related both to new types of substantive questions in a changing society and to evolving technologies for monitoring and examining more complicated phenomena than earlier, create a need for new types of statistical modelling for new types of analysis, and potentially also for new concepts of information and inference. Themes and challenges we shall work on here share the concept of the focus, the operational view that the science and context drive the most important questions which again should influence the optimal combinations of models, their analysis, and the ensuing decisions. Three such challenges are as follows. A: "breaking the wall" between areas (i) and (ii), partly leaning on recent advances inside area (iii). B: extending the current scope and catalogue of approaches and methods of relevance to area (iv). C: extending and developing new methodologies for areas (i) and (iii) for the by now frequently occurring situations where the number of measurements per individual exceeds the sample size.