author: niplav, created: 2025-04-03, modified: 2025-04-08, language: english, status: notes, importance: 9, confidence: log
.
The rough idea is to have a self-blinded randomized trial with blocking, where each block is a period of four days alternating between administration and recovery days.
It doesn't matter when a block starts, it can be any day of the week, and one can take a couple of days off from the experiment when something (e.g. a holiday) comes in the way.
I would like to keep the measurements manageable and scalable: There is a core of measurements performed in each block, but if I feel like I have slack I might decide to do more extensive measurements for one block.
If done per block, this will not impact the quality of the data, but please don't decide within a block to switch the detail of measurements.
Measurements were selected according to how much they degrade with sleep deprivation, informed by this auto-generated report using Elicit.
Ideally, we wouldn't have to re-fill the nasal spray bottles after every block, instead marking the bottle in an easily blindable way (e.g. a small dot on the bottom) for being able to figure out which bottle contains what.
Unfortunately, Claude informs me that Orexin has a half-life of ~2 weeks when stored at fridge temperatures in a liquid solution, and is best stored long-term in powder-form in the freezer—hence the plan to be careful and mix it back for each block.
The blocks reduce unexplained variability. The pause days make sure that (1) one doesn't suffer from excessive sleep deprivation, and (2) to check if there is any "catch-up sleep" that needs to happen even after taking Orexin. Investigating tolerance effects would be interesting, but I just don't think we have enough statistical power to get there, and it's more useful to focus on one-off effectiveness instead1. Tolerance effects can be examined later.
epistemic status: Doing a power analysis for the first time. Checked by Claude.
Let's say we will use a two-sample t-test. We can use statsmodels to do the heavy lifting; we want to detect a medium effect size (0.5) with a bog-standard significance level of 0.05 and a power of 0.75:
import statsmodels.stats.power
>>> statsmodels.stats.power.tt_ind_solve_power(effect_size=0.5, alpha=0.05, power=0.75, alternative='two-sided')
56.49860618876443
This means that we'll have to collect ~60 Orexin samples to detect a medium effect size; splitting it over two people means that each takes Orexin ~30 times and Placebo 30 times, with blocks of a length of four days that gives us 240 days in total (though half of those don't actually entail any work in terms of data collection).
Even demonstrating one-off effectiveness would be cool: It's often the case that people have short-term sleep deprivation, and would like to ameliorate the effects. ↩