Statistics is used every day in data analysis by particle physics collaborations: our results are entirely dependent on the application of statistical methods, from very simple ones to cutting-edge tools. If we measure the properties of a subnuclear particle with some collision data, we are doing "point estimation" - a wide topic in Statistics. When we want to assign a uncertainty to that estimate we are doing "Interval estimation" - another huge piece of Statistics literature. And if we set upper limits on the production rate of a reaction we are again doing interval estimation - only, the interval is one-sided, as we include zero as the lower bound.
Then, of course we employ goodness-of-fit measures to determine if we understand our data; we do hypothesis testing to exclude or allow a model; we use unfolding techniques to derive distributions at production level, "unsmearing" the data from the effect of the detector resolution. We apply multivariate techniques to select small signals among large backgrounds (classification) or to improve the measurement of an observable quantity (regression). And we do averages of different point estimates, include nuisance parameters in our fits through Bayesian or frequentist techniques. We estimate variances with bootstrap methods. And I could go on!
While the above should convince anybody that physicists are expert statisticians, this is not really the case. The knowledge of basic statistics of the average particle physicist has improved in the past twenty years or so, but it is still not very broad. For this reason, schools of statistics for physicists have started to be offered, to improve the knowledge of basic concepts.
But then there are higher-level challenges we face. There are subtle issues in the application of statistical methods to the analysis of particle physics data: the issue is often the ill-posed nature of the problem. Statistics allows you to derive different results from a set of data depending on who you are, because there are slightly different ways to ask the same question to your data, and details matter a lot. Advanced statistical methods also offer new opportunities to improve the sensitivity of your measurement, but mastering them is not for everybody.
In short, discussing statistical techniques should be an occupation to attend with more care by particle physicists. That is why a parallel session for that purpose is, in my humble opinion, a very good idea - and strangely enough, one that appears to be new (but please correct me if I am wrong). While there have been specific workshops on Statistics in the past, the idea of inserting a session on Statistics in a conference devoted to particle physics is a novelty.
So what are we offering in Thessaloniki this year ? Here is a quick glance at the program.
- Computing averages
- Confidence intervals for the ratio of quantities
- Mixing techniques for background modeling
- Systematic uncertainties reduction for jet-energy scale measurements
- Bayesian techniques in HEP
- Bayesian non-parametric classification
- Unfolding methods
- Unfolding in ATLAS
- Resonance searches in hadron spectra
- Higgs search tools in ATLAS and CMS
- The look-elsewhere effect in two dimensions
- Parallel computing and significance calculations
- Matrix element methods
- MVA methods in HEP
- Bootstrap techniques for anomaly detection
As you can see, there is a mix of talks on generic topics and talks on specific aspects of data analysis problems. Several of the "generic" topics are discussed by well-known experts, like Eilam Gross, Luca Lista, Sergey Gleyzer, Stefan Schmitt, Harrison Prosper (and the others please forgive me for not listing their name here). In summary, I expect a lively discussion and the participation of an interested audience... And if you have a chance to come and attend the conference, do not miss the statistics sessions! They are scheduled for the afternoons of September 1st and 2nd.