Reproducible Research

According to this article, two economists attempted to reproduce a number of economics papers that were published in top journals. They were unable to do so in most cases, even when they enlisted the original authors to help.

This result didn’t shock me even a little bit. When I wrote my MA thesis in economics I wanted to employ a particular, and rather obscure statistical technique. I couldn’t find a single book on statistics or econometrics that contained a full description of the technique, it was apparently rather specific to the sub-sub-field in which I was working.

Over a dozen papers claimed to have used it or otherwise discussed it, but zero contained an actual description of what was done to the data to bring about the result.

I finally found a proper description in a masters thesis from someone at a university in Sweden (if I recall correctly) whose adviser had apparently just happened to know what was being done to the data and who had actually taken the time to describe it. The thesis was never peer-reviewed (although the student had apparently graduated successfully, so I felt comfortable relying on it). So while I still had to implement it myself and verify my results, at least I knew where to start.

The situation is even worse given that it is fairly rare (in my admittedly limited experience) for authors in economics (or other social sciences) to publish their original datasets and (perhaps even more importantly) the code that they ran to do their analyses. I suspect that many couldn’t even if they wanted to due to a reliance on tools like Excel and SPSS that do not lend themselves to replicability without significant extra effort.

This is not to say that economists are evil or that there is some kind of conspiracy (although some are evil, and there are almost certainly “conspiracies”, the replicability problem just isn’t evidence of it).

Part of the problem, I think, is that many, or even most, economists never learn about tools they could use to do a better job at promoting replicability. Version control (Git, Subversion) and tools like GitHub or self-hosted alternatives (why don’t universities run these for their faculty?) are a great start. Using proper statistical languages and doing 100% of analysis using code, not “interactive mode” would help as well.

However, the real key, in my opinion, is for people to get comfortable working out in the open. I publicly publish virtually every line of non-trivial code I write. A lot of it is complete garbage, but I publish it because there is simply no reason not to do so. I’m writing my computer science thesis entirely in the open, from the very first paragraph.

I do realize, of course, that the stakes for me are very, very low. Academia is not my career, so I don’t worry about getting “scooped”, or about being attacked by a colleague with a vendetta. But that just means that some people might want to employ a self-imposed embargo before releasing their work. Wait until your grant runs out, or until the paper is actually published, and then put everything online (and no, simply dumping a PDF on arXiv doesn’t count, do it right). I honestly believe that every field would be better off if this were the norm rather than the happy exception.

As an aside, for anyone interested, Roger Peng, a biostatistician at Johns Hopkins, has an excellent Coursera course on reproducible research. I watched some of the lectures and it seemed like a great course on an important topic. As a bonus, Dr. Peng is a fantastic lecturer, his courses on the R programming language are also top notch (and accessible enough for reasonably bright social science students).

Image credit: Janneke Staaks