All of snul's Comments + Replies

Case study: Melatonin

I attempted to use melatonin on myself in a pseudo-study and wrote about it here: http://lesswrong.com/r/discussion/lw/ngj/open_thread_april_4_april_10_2016/d7s5

Open Thread April 4 - April 10, 2016

Following gwern's post about Melatonin, I did a self-trial of melatonin to see if it improved my sleep.

Objective: To see whether I (myself only) should take melatonin regularly or not. Therefore: Have a clinically significant decrease in sleep time or increase in daytime awakeness. Clinically significant=statistically different & meaningful (i.e. 1 minute difference is not worth the effort, for example)

Note that at baseline, I never have trouble falling asleep, and I never have trouble staying asleep. Melatonin is theoretically advantageous in inducing... (read more)

0[anonymous]5yGood job on the study. I wouldn't call that a low n. I'm guessing you used a power analysis such as this one [http://clincalc.com/Stats/SampleSize.aspx] to calculate the sample size prior to the study, but the results are clearly wildly different from whatever your initial projections were, so the n you calculated prior to the study doesn't tell you all that much. Just eyeballing the covariance in your graph of the cumulative distribution (probability is usually listed on the y axis, btw), you've got a very strong result there. I do think intention-to-treat was the wrong system. I would have excluded all nights where you didn't want to use melatonin from the study, and randomly assigned nights where you were considering melatonin to either treatment or baseline. It looks like correcting for that would have given you even stronger results, however. But I think you went way past the point where more nights would have been of much use. Melatonin just doesn't appear to make much difference for you.