From xkcd: Kinda related to this and this ..read more

Just got a weird phone call from two people who claimed to be from the New York Times, asking me very vague questions. They said someone recommended my name to them, but they wouldn’t say who recommended them. They asked me what my research was about, and I said I do a lot of research so it would help if they could tell me what their story was about. They said they couldn’t tell me. I recommended they go to my webpage and look at my published and unpublished papers and then if they had any questions they could ask me something specific. Or they could send me an email. It was a very weird conve ..read more

Colleagues of mine here at Novartis – Björn Holzhauer, Lukas Widmer & Andrew Bean – and myself have recently published the new website on “Applied modelling in drug development via brms”. The idea is to present real-life case studies from drug development and demonstrate how these can be solved using the amazing R package brms by Paul Bürkner (short for Bayesian regression models using Stan). brms lets you specify your model using the R model formula syntax you may be familiar with so that you do not need to write your own Stan code – while offering a surprising number of features in a sup ..read more

Kuang Xu writes: I’m a professor at the business school at Stanford working on operations research and statistics. Recently, I shared one of our new preprints with a friend who pointed out some of your blog posts that seem to be talking about some related phenomenon. In particular, our paper studies how, by using adaptive control, the states of a processing system are effected in such a way that congestion no longer “correlates” with the underlying slowdown of services. You mentioned in the blog where you wonder if there’s some formal treatment of this phenomenon where control removes correla ..read more

This is Jessica. Human subjects experiments are starting to appear in machine learning venues. For example, ICML, one of the big ML conferences, accepted a few papers on quantifying prediction uncertainty with user studies.   Overall, I’m glad to see this. Theoretically rigorous uncertainty approaches that provide calibration or coverage guarantees may not be worth the effort (e.g., in terms of the model retraining or holding out of additional data that they require) if providing that information doesn’t impact human decisions, or leads to under- or overreliance. We know from decades ..read more

I was reminded of the above line (it’s from a Leonard Cohen song) after reading something stupid on the internet regarding some technical issue, and various people were trying to place the dispute in the context of a so-called academic “war.” I won’t get into the details because there are a million such arguments on the internet, and here I want to focus on a general problem in communication that this example illustrates. I’ve been involved in lots of academic disagreements over the years, and I pretty much never think it’s a good idea to frame any of them as a “war,” even as a joke. Indeed, I ..read more

Word on the street I often hear people say things like “We do MCMC to get a full representation of the posterior.” Their intuition seems to be that MCMC is going to take a set of draws that covers the posterior in some way. Misleading textbook examples I’m just as guilty of this as everyone else. The textbook examples are in 1D or 2D. We throw darts at a unit square and keep the ones in an inscribed circle to calculate pi, for example, or we sample independently form a 1D distribution and fill in a histogram (in the limit of infinitely many draws and infinitesimal bins, the histogram approache ..read more

From Ray Monk’s biography of Bertrand Russell: Though the Russells were not especially wealthy, they employed—as was common in Britain until after the Second World War—a number of servants: a cook, a housemaid, a gardener, a chauffeur and a nanny. Arguably this is not so much different than modern society: even if we who live in comfortable circumstances do not employ personal servants, we still benefit from the labor of thousands of people working in farms, factories, and everything in between. What struck me about the above story regarding Russell is not so much that he had all these serva ..read more

Greg Sargent writes: I disagree with Sargent’s statement that “The reason Trump regularly tells lies that are very easy to debunk . . . is to assert the power to say what truth is.” I think it’s simpler than that. People like to say things that make them look better, and this is easier to do if you’re not constrained by the truth. The big question is not so much why someone who has lied a lot in the past keeps lying—people typically keep doing with what’s worked for them before—but rather why so many of his supporters don’t seem to mind. And there I’d like to draw a connection to something ..read more

Elden Griggs writes: I have been grading a homework problem where the students have come to two different conclusions. The question was: You want to gather data to determine which of two students is a better basketball shooter. You plan to have each student take N shots and then compare their shooting percentages. Roughly how large does N have to be for you to have a good chance of distinguishing a 30% shooter from a 40% shooter? Essentially, there were two sets of student answers. The first set simulated and found the answer to be around 120-130 depending how they rounded (generally they w ..read more