Can you show us that again please? - BioMed Central blog
"As an undergraduate at Cornell in 1985, I looked for a research problem that combined my interests in genetics, evolution, and behavior. Kyriacou and Hall had recently reported that the period gene, which regulates circadian rhythms, also controlled a rhythm of fruit fly courtship song and that evolution of period explained a species difference in this courtship song rhythm. This seemed a promising avenue, so I surveyed molecular variation at the period locus in Drosophila species, hoping to correlate this genetic variation with variation in the song rhythm. I hand-scored the fly songs, attempting to find the song rhythms. I couldn’t find them. I figured 'Heck. I’m an undergraduate. I don’t have much data. I probably screwed up.' I let the problem rest. About six years ago, though, I decided to try again. We built a device for collecting courtship song and wrote software to score the songs automatically. We looked for the rhythms, but we couldn’t find them. This observation kept nagging at me, however. So, I set out to reproduce Kyriacou and Hall’s experiment as closely as possible. Still, I couldn’t find the rhythms. Finally, when I analysed the data precisely as Kyriacou and Hall had—binning the time series data and scoring only the strongest periodic signal, patterns similar to the rhythms popped out of this analysis. however, almost none of the signals I found reached any sensible level of statistical significance. While I was able to “replicate” Kyriacou and Hall’s song rhythms, they were artefacts of data binning and did not result from a real biological phenomenon. Recently in the scientific press, there has been considerable discussion of scientific fraud and of research findings that cannot be replicated. While outright fraud is probably rare, a much more widespread problem is the publication of results which, like those of Kyriacou and Hall, were got at through honest scientific investigation but that cannot be replicated. How can we minimize publication of results that are not reproducible? I have one simple prescription. Authors should show their data ..."