Empirical Sleep Time

by rlpowell1 min read24th Jun 201231 comments


Personal Blog

I'm thinking that it should be possible to decide when to sleep based on reduced performance.

Can anyone suggest a tool for that purpose?  Perhaps some reaction time testing software?

I guess I would have to track myself during the day to make a baseline, which is fine.

But without some sort of test I end up staying up way pass effectiveness, which is a waste of my time.


31 comments, sorted by Highlighting new comments since Today at 12:56 AM
New Comment

Good and bad sleep accumulate over days. Thus moment by moment measures of productivity are of limited use and any experiment has to compare extended periods of time, a week at the least, probably better a month. The very widely held consensus in the field is that you should go to sleep at the same time every night and wake up without an alarm. This should be the control for any experiment.

If only I had that option. :) If you know how to explain this to my 8 month olds, please do let me know. :)

I appreciate the info about the ideal there; I'm going to keep it in mind. But it's simply not reasonable for me right now.

My focus and attention are much better late at night, so I tend to stay up as late as I can stand, to get more done. Unfortunately, with babies around, both when I wake up in the morning and when in the evening I can start focusing on the things I want to work on are entirely random. Hence wanting to get as much effective time out of the late night as I can, hence the original question.

Obviously, performance does degrade if one stays up long enough; I simply want to have something better than "I'm passing out in my chair" as a hint that I should give up and go to bed.


I'm thinking that it should be possible to decide when to sleep based on reduced performance.

It certainly is!

Can anyone suggest a tool for that purpose? Perhaps some reaction time testing software?

Double N-back seems to be precisely what you're looking for. It effectively tests how long a sequence of information you can keep in your mind, and how accurate you can be about recalling the sequence. Being tired impacts your DNB performance, and is a marker for the kind of "reduced performance" you're looking to avoid.

Accounting for training effects aside, the way you'd work this is to test yourself on a DNB regime every hour or so. When your results start to drop, you entertain the hypothesis that you're tired; if the next hour is further down still, you decide you're tired to the point of reduced performance and you go to sleep.

(That is a rule of thumb calculation; you can get Bayesian if you want)

That's exactly the sort of thing I had in mind, thank you! I'll try it.


Can anyone suggest a tool for that purpose? Perhaps some reaction time testing software?

I used something called a Psychomotor Vigilance Task when I was doing my sleep deprivation experiments. Essentially, you spend ten minutes watching a screen, reacting when it starts counting up. This measures how frequently your attention lapses (since even at high levels of sleep deprivation, it's possible to accomplish complicated tasks, so long as they're short duration). I coded my own in Ruby Shoes, and can send you the code if that'd be helpful.

While it's useful at measuring sleep deprivation, I don't think it's particularly useful at measuring performance around normal levels. What you're interested in there is not so much "how easily do I get bored?" but "how creative am I?", and the PVT is really unfun and boring to do.

If you don't get enough sleep one night, you'll "rebound" and sleep in longer the next time you get a chance to do so to make up your "sleep debt." If you log your sleep every night with a ZEO (or similar tool) you can get a pretty good estimate of how much sleep you need just by looking at how little sleep you can get without a rebound on the weekend.

Performance can be a tricky issue, because short-term sleep deprivation can improve your mood and alertness, whereas short term sleep deprivation does the opposite. You may find that you do better on a test with less sleep, if you had been well rested up until the night with restricted sleep.

In general the effects of something that happens on one day carrying over to the next is a difficult problem to deal with in self experimentation. I would recommend trying to get more or less sleep than usual in about the same amount for at least a week before drawing any inferences about performance at that sleep duration.

In general the effects of something that happens on one day carrying over to the next is a difficult problem to deal with in self experimentation.

It's not really difficult: you solve it by using multi-day blocks. In my self-experiments, I randomize over days or weeks without problem.

That's what I was trying to get at in my post, but I wasn't being very clear.

However, if you randomize over multi-day blocks your data points go down. Once you're randomizing whole weeks and looking at parameters with a high standard deviation, it might take impractically long to collect meaningful data.

I'm starting to think that long several week blocks are absolutely necessary for drug or lifestyle self-experimentation, because the effects once tolerance is developed are usually different from naive effects. For a drug or lifestyle change that's intended to improve quality of life or performance, the effect of long term steady use is usually what you're interested in.

However, if you randomize over multi-day blocks your data points go down. Once you're randomizing whole weeks and looking at parameters with a high standard deviation, it might take impractically long to collect meaningful data.

Que? If I'm randomizing week-blocks and I'm measuring my sleep data each night, and I have 7*4 nights on and 7*4 nights off, don't I have as many datapoints as if I randomized 56 individual nights on and off? I'm not testing solely at the beginning or end of each block. Even if it takes a full 6 days to reach steady-state, I still get signal from day 7.

The data doesn't become representative of what you're trying to test (long term use), until after it's stabilized or tolerance is developed.

For example, imagine you're testing a stimulant which keeps you from sleeping the first few days, but eventually helps you focus better and sleep normal. You can't use the sleep deprived data from the beginning of the cycle if your goal is to identify the effects of using it long term.

I guess it's obvious, but I was just pointing out that it takes longer to do a self experiment on something that has effects which change gradually over time vs something that can be assumed to stabilize quickly.

Is your problem determining when to go to sleep, or actually making yourself go to sleep when you think you should for optimal effectiveness?

Given external evidence that my performance is reduced, I don't think getting myself to go to sleep will be a significant problem. My issue is usually the belief that I'm still getting useful stuff done, and actively resisting sleep on that basis; evidence to the contrary is something I think I would treat as real data.

[-][anonymous]8y 0

Wow, I'm appalled by how much I've become dependant on external supports: the first time I took the remember-nonsense-words and the remember-digit-sequences tests, I instinctively wanted to reach for my pencil (because that's what I would do ‘in real life’ to remember a sequence of digits or similar), and I felt that having to tell my hand not to do that was quite distracting.

This would have been better in an open thread.

If you mean I should have put it in the "main" tab rather than the "discussion" tab: I thought that was, essentially, for coherent essays, rather than Q-and-A. If that's not what you meant, please explain?


Each month there is a post in the discussion section called "Open thread for May 2012" (or equivalent). He means you could put it as a comment there.

I have no problem with this as a discussion post myself.

There exists a monthly Discussion thread to which people can post as comments various things which aren't worth a whole post.

I would think there would be too many confounding variables, and the best you could do is find some testable proxy for performance.

You could just buy a Zeo.

Unless I'm misunderstanding, a lot of what the Zeo does is select a wake time that matches one's sleep cycle. I have twin infants; the time(s) at which I am awoken are entirely out of my control. :D Also, I don't use an alarm, and haven't in many years; as such I awake at what I assume is the right point of my sleep cycle (when allowed to wake naturally, which doesn't much happen anymore).

If I'm misunderstanding, feel free to elaborate; the copy on Amazon at least is pretty inspecific.

Generally speaking, though, I don't have a serious problem with sleep debt or effective sleep or anything; my main sleep problem is staying up so late that I'm ineffective and hence wasting my time, but I do generally get enough sleep to compensate so that I'm OK in the morning.


I had a Zeo for awhile. It does a pretty good job, but it was disturbing how locked-in your data was. There are smartphone apps that will also gather comparable data for much less:


Also, if you snore, then blood oxygenation level is not a bad proxy for sleep, and clinical quality devices are available on ebay for much less than a new Zeo.

ZEO isn't locked in at all anymore- they have a decrypted firmware you can download, an open source library for interfacing the data format, as well as instructions online to interface the ZEO directly via a serial port if you want realtime data.

You can now use the ZEO without even ever uploading the data to their website.

Locked in?

Also, I haven't see any reliable information to the effect that accelerometer-based inferences are comparable to the Zeo's electrode recording.

[-][anonymous]9y 2

I'm confused by this thread. A Zeo isn't going to tell you when you ought to go to sleep.

No, but it will tell you if you aren't getting enough in general, which is more relevant since sleep debt can build up over several days.

[-][anonymous]9y 0

It's not what OP is asking about.

If the OP is interested in performance as it relates to sleep, this is probably an interesting area to explore. Metaphorically, if I see someone struggling to pick apples from a tree here, I might direct them to the low-hanging oranges on the tree there.

That said, I also did answer the OP's question up-thread.

[-][anonymous]9y 0

Yeah, I guess Zeo would be better, but I dunno whether it is so much better as to justify the larger price (two orders of magnitude more expensive than Sleep Cycle).

Yes, locked in. When I was using the Zeo, all of your data lived on their flash based site, and there was no obvious way to get it out and into your own analytic tools.

Also, for purposes of managing sleep hygiene, you don't need highly accurate records. Moderately accurate data will benefit most individuals with imperfect sleep hygiene.

Perhaps it's a new functionality, but their site currently offers data export to CSV.

They've had CSV export ever since I began using it, and I've never had major problems with that - it's what I use for my statistics in R. Since around when I got one, they even have firmware that supports sending out the raw data second by second over a serial cable (which I was thinking about using for meditation).

So I figure either stcredzero used Zeo a long time ago, or he simply didn't know about the options.