Adventures in Onionism

More research on how research is a crock. Star Wars: The Empirics Strike Back , a working paper from the IZA website. Basically, they show the distribution of published p-values is funny-shaped, and what you'd expect if people were specification-shopping to get a barely significant p-value. You know, publication bias. [h/t WB Impact Blog]

An empiricist making policy decisions instead of an ideologue? I know, I'm as surprised as you are, but that's apparently what Washington state got in the guy they picked to advise them on their experiment with legal weed . Economist Hulk on Twitter . Most of the expenses from the wars in Iraq and Afghanistan are yet to come. Let me say that again: Yet to Come .

Most published research results are false . Publication bias is a serious problem. Nobody posts their data online. Replications aren't sexy enough to get published. Here's a blog about some of that: Political Science Replication . UPDATE: All the smart people admit everything is screwed up. Today Blattman responded to a series by 10 CEGA people about this same sort of thing.

This article by Mark Bittman has pretty interesting stuff on causality in medicine, given that it's impossible to run a randomized trial where you randomize a lifetime of smoking or eating lots of sugar. Teaching is exhausting, but fun. This paper by Nunn & Puga from REStat last year is my favorite article I've taught so far this semester. They claim ruggedness is bad for GDP in most places of the world for obvious reasons (transport costs), but it's good for GDP in Africa because it reduced the slave trade, which is even worse. Really no fancy econometrics, just a simple design and lots of robustness checks. You don't want to run a regression of GDP on ruggedness because omitted variables will obviously bias your estimate, but it's perhaps less bad to run one on ruggedness, an Africa indicator, and the interaction of ruggedness and the Africa indicator. The coefficient on ruggedness is still biased, but the coefficient on the interaction is only biased if t...

Or rather, make no difference. I was looking for bad observational statistics articles to share with my stats class which starts next week, and accidentally found stuff that's apparently good: RCTs about shoes . Motion control shoes may control motion, but they may injure you, or at least not help you. N=81 said they injured women, N=~1300 said they make no significant difference.

Just finished reading Eat and Run: My Unlikely Journey to Ultramarathon Greatness by Scott Jurek (with Steve Friedman). While I'm happy that the book has motivated me to run a little (lot) more consistently, I mostly didn't like it, for two reasons. One, the writing was tired and cliched. Two, there were too many pseudo-scientific and psuedo-spiritual claims about the source of his running prowess (e.g. homeopathy, cleansing toxins, etc.). For good measure they mentioned one of the many horrible statistical studies showing that sitting makes people die early (tangent you've heard before: just because you control for smoking and other aspects of lifestyle doesn't mean you get at causality. If there is anything in the world that you did not include in your analysis that is correlated with both sitting and life expectancy, then your estimate of the effect of sitting on life expectancy is incorrect. And if there are plenty of things that you can measure that are correlated ...

Given my research idea (and my film snobbery), this article about the box office success of Paul Blart: Mall Cop is pretty interesting. This poll is the best poll ever. If I were a big nerd, I'd tell you that if you assume there's exactly a 50% chance that international voters choose heads, then the odds of 1857 of the 3439 international voters choosing heads are exactly 3439!/(1857!1582!)*.5^1857*.5^1582. If were an even bigger nerd I'd simplify that expression for you using Matlab, because it seems it would take me more than 30 seconds to figure out how to get google, Excel, or Stata to crunch numbers that large/small for me. Oh well.