Image credit: Enes Evren.
I’ve been trying to quantify, understand, and improve my sleep for many years now, and in the following, I want to share with you my longest and most in-depth self-study yet.
So let’s take a look at what effects caffeine, alcohol, and exercise have had on my sleep.
I started this analysis in mid-2018 when I wrote about the results of tracking and modeling my blood caffeine concentration for a month.
How Much Coffee is Too Much?
Tracking My Daily Caffeine Consumptionmedium.com
Towards the end of 2018, I shared the results of a more detailed experiment, covering around three months of sleep and related data and attempting to find correlations.
While providing some interesting insights, the data was far from statistically significant. But since then, I have kept up the tracking, measuring my sleep data each night with my Oura ring, and logging every caffeinated beverage I consumed, my workouts, and my daily alcohol consumption.
In total I accumulated 632 days of data, spanning from September 2018 to December 2020 (with two unfortunate small gaps in the data, more on that below). This is a detailed analysis of this data and what I learned from it about my sleep and my stress levels, as well as how my habits and behaviors impact them.
First, we will take a look at my nightly heart rate and heart rate variability (HRV), which are key indicators for recovery, stress response, and overall health and performance.
Then we will dive a bit deeper into the science of sleep, its individual phases, and look at how different factors influence my sleep phases.
Finally, I want to draw some conclusions on what I managed to learn from this experiment.
But before we dive into the results, a quick word on how I collected and modeled my data (feel free to skip this part if you just care about the results).
Data Collection and Modeling
My experimental journey began way back in 2015 when – driven by my lasting obsession with coffee (and a slight worry that I was drinking too much of it) – I decided to log every single caffeinated beverage I consumed using the simple app rTracker.
Later I also started to use rTracker to log my alcohol consumption. Each morning I would assign a label of “None”, “A bit” (roughly a glass of wine or beer), “Quite a bit” (on the order of three to five beers), or “A lot” (anything beyond) to the previous day.
Sleep was even simpler to track. Ever since their original Kickstarter campaign, I’ve been a big supporter of the Oura ring, a sleep tracking device I wear constantly. So getting my sleep data was as simple as downloading a CSV file containing the well over 50 metrics the ring tracks through their Oura Cloud service.
And while Oura’s primary focus is on sleep, it also measures activity during the day, and the app can be used to log additional workouts and activities.
Armed with all this raw data I was almost ready to start my analysis. Only one piece I was interested in was missing: if I wanted to find out how caffeine affects my sleep, I should first know how much caffeine was in my blood by the time I go to bed. So I had to create a model that could take my log of caffeinated drinks and estimate the actual caffeine levels in my body based on it.
Leading up to the first article mentioned above, I made a model that could roughly tell me the amount of caffeine I had coursing through my bloodstream at any point in time.
For the full details of the model (as well as more info on how exactly I tracked all the data and what tools I used) I refer you to my previous articles, or the code itself, but here is the core caffeine model in a nutshell.
My caffeine logs were split into the main types of caffeinated drinks I consumed – filter coffee, espresso, cold brew, tea (cup), tea (bottle) – and an average caffeine amount for each, as well as an “other” category, for which I could specify a unique amount of caffeine in mg.
In reality, caffeine concentration varies highly, even between different coffee beans, so that assuming a single value for e.g. my usual filter coffee made with 13g of coffee beans (let alone a coffee I’d get at a coffee shop) is a pretty strong simplification. But I believe that the resulting model is still insightful, and simplicity is key for actually turning the tracking into a habit that sticks. I also played around with a range of different values, and while the exact caffeine values in my blood change, all the results I will talk about below still showed the same relations across the values I tested.
To actually model caffeine intake, I assumed that the rate at which caffeine enters my blood takes the form of a Gaussian distribution that peaks 60 minutes after consuming a drink (which I took to happen in a single moment in time in my log, usually the time when I started drinking).
Further, I assumed that the half-life time of coffee in my blood was four hours. This seems to be a reasonable estimate, maybe even a touch conservative given that I know from DNA analysis that I’m a fast caffeine metabolizer.
Feeding my raw logs (in the form of CSV files) into my model, which I coded up in Python, I could then get a reasonable estimate for how much caffeine I had in my blood at any point in time.
While the individual days can vary widely (and we’ll look more into this later when relating it to my sleep), below you can see what this looks like averaged over all 638 days in my dataset.
The amount of caffeine in my blood averaged over all days. The dashed lines show the caffeine peaks, the dotted lines show the caffeine at my average bedtime.
You can see that I tend to get up between 8:30 a.m. and 9:30 a.m. (a bit later on the weekends), immediately have my first coffee, sometimes followed by a second coffee between 10:30 a.m. and noon. Then there’s a little dip as my body gets rid of some of the caffeine while I have lunch (and maybe even a short nap), but in the afternoon I usually top up with another cup or two, this time in a slightly less regular fashion than in the morning.
On the average weekday, my caffeine levels peak at just over 120mg at around 16:40 and then decay into the evening and night, but still being at over 40mg of caffeine by the time I go to bed around midnight. Even by the time I wake up, I usually still have the equivalent of a quarter espresso in my blood before the cycle begins again.
Half-life times can mislead our intuition a bit, but seeing it plotted like this makes it clear just how long the caffeine stays in your blood. And that’s only for my fairly average consumption of three to four cups of coffee a day, and a (probably) shorter than average half-life time given my fortunate genes. For many people, this might look much worse.
Just like caffeine metabolism, how much caffeine affects us also depends on the individual. I can’t speak for everyone (although I hope my results give some good general indications), but my N-of-1 experiment should at least give me some concrete insights about myself.
So should I be worried? Is this ruining my sleep? Do I have to give up my beloved coffee? Let’s find out.
Resting HR and HRV – Indicators of Stress and Recovery
The heart is one of our most critical organs. It’s responsible for pumping blood around our body and supplying it with oxygen and other important nutrients.
For a long time, the heart, rather than the brain, was even seen as the center of the body and the seat of emotion. And more and more modern research is suggesting that this was not completely off, that there is at least some kind of two-way communication between heart and brain.
So the heart is clearly important for physical, as well as mental wellbeing.
The most common and well-known metric related to the workings of the heart is the heart rate (HR) – a simple counter of how many times our heart beats per minute.
Heart rate variability (HRV) is a less well-known quantity, but it’s quickly gaining popularity as a key quantity to keep an eye on if you’re interested in your health and performance.
Our heart does not tick like clockwork. A heart rate of 60 bpm does not mean there is a beat exactly once every second.
Instead, the duration between consecutive beats varies. Our heart adapts to the slightest nuances in our environment and the ever-changing requirements from our body. It also speeds up with the in-breath and slows down with the out-breath.
At least that’s the ideal scenario.
The more rested and relaxed we are, the stronger this response. But if we are stressed, fatigued, or otherwise in a sub-optimal state, our body’s response gets dampened. It’s a sign that our sympathetic nervous system, responsible for our fight-or-flight response, is overactive.
This makes HRV a great metric to measure how well-rested and ready for peak performance our body is.
In short, the higher our HRV, the stronger our adaptability to stress and the environment.
I have recently spent a considerable amount of time and focus on trying to master and increase my HRV, following a ten-week program of twice-daily breathing exercises to deliberately increase my HRV and reduce stress and anxiety. The results of this have been quite profound, and deserve a separate article which I’m planning to write soon.
For the sake of our analysis here, it’s enough that we agree that higher HRV is better.
The Oura ring measures and reports three different quantities related to the function of the heart during our sleep: HRV, as well as the average and lowest heart rate during the night.
Looking at the plots below, we can see that these quantities are quite strongly correlated.
Resting heart rate and heart rate variability are closely correlated, with lower heart rates corresponding to higher (i.e. better) HRV values. (All images which follow courtesy of the author.)
Nights during which my heart rate was lower, were also nights with higher HRV. Similarly, average and lowest HR follow each other quite closely, with the average usually being around 5–7 beats higher than the minimum.
Having established that HRV and resting HR are good indicators of how well-rested we are, we can use it as a proxy for overall sleep quality and stress to see how different factors impact my sleep.
The plots below show how my heart rate and HRV vary with the amount of caffeine in my blood at the time I went to bed.
Impact of caffeine in my blood at bedtime on heart rate and heart rate variability during sleep. Caffeine appears to increase HR and decrease HRV, but mostly above a certain minimum threshold.
As one might expect, we see that caffeine seems to have a slightly negative impact on my rest, increasing my heart rate and decreasing my heart rate variability.
However, assuming a linear relationship between these variables is definitely a strong oversimplification, and a closer look suggests that the negative effect might be largely dominated by outlier days with particularly high caffeine, over around 80–100mg.
I have to admit that I’ve not done anything close to a thorough literature review, but from a quick look at several papers that study caffeine’s negative impact on sleep, I noticed that many of them only look at the higher end of caffeine amounts (sometimes 300 mg or more) administered in a rather unrealistic single large dose.
I don’t want to make any unfounded claims or accusations, especially since I haven’t looked too deep into the literature, but it does make me wonder if many of the studies reporting negative effects of caffeine on sleep ignore a range of low to moderate caffeine and more realistic consumption where these effects are actually all but negligible. At least that seems to be the case for me.
Overall, my data seems to suggest that as long as I don’t overdo it with the caffeine, my coffee habit doesn’t affect my sleep too badly.
Physical Activity
The next variable I want to look at is how active I was during the day. For this, I use Oura’s Activity Burn metric, which estimates the number of calories I burned through activity, based on my motion and physical data, as well as exercises I logged in the app (e.g. Crossfit workouts, during which I’m not wearing the ring).
Looking at the data, we can observe a slightly positive impact of exercise on my sleep quality.
Impact of activity level on heart rate and heart rate variability during sleep. Exercise appears to decrease HR and increase HRV.
This was actually a little bit surprising to me.
I didn’t doubt that exercise had a positive effect on my sleep over the long term. But the night immediately following the exercise, which is what’s shown in these plots, I would have expected a slightly increased resting heart rate and decreased HRV since my body had been under stress and now had to recover.
One possible explanation could be that exercise reduces stress, and on days on which I exercised a lot, I was considerably more relaxed. Maybe this mental effect and lower stress hormones more than made up for my tired body.
Or it might be that I exercised more on days on which I already felt particularly good and well-rested, and this carried on into the night.
One variable that did however behave as expected was the restlessness of my sleep.
Exercise appears to lead to more restless sleep the same night.
In accordance with my subjective perception, more exercise did seem to lead to more restless sleep (although unfortunately, I’m not sure what unit Oura’s “Restless Sleep” metric is in, or what exactly it measures, so I might also be misinterpreting this a bit).
Often after a tough workout, especially if it was later in the day, I do feel like I’m tossing in bed a bit more than usual, and the data seems to confirm this suspicion.
Despite this, exercise for me seems to have a positive effect not just over the long term, but immediately.
Next, onto a variable whose effects are less subtle and difficult to interpret: alcohol.
Plotting my heart metrics against the amount of alcohol I consumed shows without a shadow of a doubt that alcohol is bad for my sleep and recovery. Really bad.
Impact of alcohol consumption on heart rate and heart rate variability during sleep. Alcohol considerably increases my heart rate and decreases my HRV.
While a bit of alcohol doesn’t seem to make too much of a difference (but is still definitely noticeable in the data), quite a bit or even a lot of alcohol dramatically increases my heart rate and lowers my HRV.
Oura also provides an overall readiness score spanning from 0 to 100, which is a compound calculated from many different variables that indicates just how ready and well-rested you are.
Impact of alcohol consumption on overall readiness. Alcohol drastically decreases readiness.
Here the effect is even more drastic. Beyond a little bit of alcohol, the next day’s readiness score drops considerably.
Even my breathing is affected by alcohol.
Impact of alcohol consumption on respiratory rate.
While the effect is not as strong as on the other variables, my breathing does appear to slightly speed up when I had more than one drink.
So while many of the other findings here are subtle and it’s not always obvious that an effect is not actually the result of a third underlying hidden variable, I’m pretty confident to – once again and unsurprisingly– conclude that alcohol is really bad for your sleep.
For this reason, I have also excluded days with “Quite a bit” or “A lot” of alcohol from all scatter plots presented in this article, such as the activity and caffeine plots above. As I’ve shown in my previous article, not doing this can easily lead to some strange (and faulty) conclusions, like more caffeine appearing to lower my heart rate.
Looking at the Data Over Time
So far we have only looked at each day in isolation. However, this is another oversimplification for the sake of analysis. In reality, most factors probably have long term correlations and effects on each other. A single particularly good or bad night does not stand in isolation but affects the following days and nights as well.
Fully unraveling and understanding the complicated relationships between these interlinked variables over time is pretty much impossible with the data on hand.
But a simple and hopefully still enlightening first step is to look at moving averages over time.
Below is my heart rate and HRV data for the entire range from September 2018 to December 2020, averaged with a moving window of 14 days to smooth out day-to-day variations and focus on longer-term trends.
Moving average of my heart rate and HRV, with a window size of 14 days.
Overall, my heart metrics seem to have gotten worse over the timeframe I’ve been tracking, with a particularly noticeable drop in 2020.
The extreme drop in HRV (and increase in HR) at the very end of 2020 is probably a bit of an outlier though, caused by me traveling back from Japan to Germany for the holidays (in itself quite a stressful experience), the ensuing jet lag, as well as the overeating of my grandma’s Christmas cookies and overall consuming more sugar over just a few days than in the entire rest of the year combined.
There are two unfortunate gaps in my data. The first is due to me not tracking caffeine and alcohol for a while in mid-2019, whereas the second one is due to my Oura ring breaking. (And I was too busy/lazy to contact Oura straight away. Once I told them, they immediately sent me a free replacement ring. Great customer service!)
Especially the gap from late December 2019 to May 2020 is particularly unfortunate. In May I self-published my book. The months leading up to the publication, exactly the months, unfortunately, missing in the data, were some of the most intense I ever experienced, packed with writing, editing, designing, and marketing the book on a very tight schedule.
But while they were intense, it was a very positive and rewarding intensity, much more eustress than distress. I poured a lot of energy and focus into it, but also got a lot of energy and a sense of purpose and meaning back from it. I would have loved to see my metrics for that time.
Once the book was out in May, it was almost like I faced a kind of void, a feeling of “now what?” which I’m still only slowly recovering from.
I’m wondering if part of my seemingly increasing stress that’s apparent in the data is a result of this. With the data gap, it’s just speculation, but it looks like my 2020 actually started out pretty well and only went into a steady decline after the book launch in May.
Ironically, but not too surprisingly, removing the acute eustress of this meaningful project might have led to a slow creep of distress. We talk about this in our book, in the form of the ancient concept of Noble Leisure. The best form of leisure is often very active and infused with a strong sense of meaning. Remove that, and you create negative stress.
In addition to that, I am also aware of some other stressors in my life during this time period, about which I might write in a future article. And of course, there was just the general stress and anxiety that almost everyone faced in 2020, either because of direct experience, or at least second-hand through the never-ending torrent of bad news.
However, on this last point, Oura recently released their own internal study of pandemic trends titled “What if Social Distancing is Good For Our Health?” that shows some surprising trends.
What they found is that on average their users actually showed lower resting heart rates in 2020 compared to previous years, probably as a result of people being able to choose their sleeping schedule more flexibly, and more in line with their own chronotype.
They also found that, with social distancing in place and most people staying home more, sleep times became longer and more stable over the week, all factors contributing to better sleep and improved heart metrics.
I guess I’m a bit of an outlier in that data. I have always been quite fortunate to make my own sleep schedule, possibly even more so in the past than in the latter half of 2020, making me maybe even more of an anomaly.
One other potential contributor to my overall decline in HRV and an increase in resting HR might be found in the following plot.
Moving average of my activity burn, with a window size of 14 days.
Over the two years that I’ve been tracking my data, my level of exercise seems to have steadily declined.
2018 was definitely my most active Crossfit year, when I was in the gym 4 or 5 times a week. This dropped to probably around 2 or 3 times a week in 2019, and essentially to zero in 2020 (although I did run a bit more frequently again).
I also used to commute to work by bike, another more or less daily activity that pretty much vanished for the part of 2020 covered by my data.
Along with my decline in exercise level and increased HRV, Oura’s aggregate restfulness score dropped from fluctuating around a very good 90 until late 2019, to a still solid but less ideal 80 in 2020.
Moving average of my Oura ring’s restfulness score, with a window size of 14 days.
The strong and sudden drop in restfulness at the end of 2019 might be the result of a particularly large number of parties and gatherings around that time. During that time I was performing on average once or twice a week at events and clubs, which is fun, but obviously not great for my restfulness...
Weekly Trends
Inspired by the above mentioned Oura study, I decided to also look at how my values vary over the average week.
The results turned out almost comical, with a consistent increase in stress markers over the week.
Heart rate and HRV across weekdays. Values are for the previous night, e.g. “Monday” means the night from Sunday to Monday.
I start the week out very relaxed, with my resting heart rate low and my HRV high, and then go into a steady decline throughout the week, apparently getting more and more stressed until Sunday, when the cycle seemingly resets in the night from Sunday to Monday.
One thing to note is that the variance is significantly increased on the weekend. This is not surprising, since those are the days I’m most likely to go out and have a few drinks, which also explains why the stress markers keep increasing on the weekend.
Removing all days with “Quite a bit” or “A lot” of alcohol shows a slightly different picture.
Heart rate and HRV across weekdays, considering only days with no or little alcohol.
Now stress markers peak on Friday, or rather in the night from Thursday to Friday, and then start dipping again over the weekend. This aligns quite nicely with the average presented in the Oura study.
Saturday to Sunday still has a slight decrease in HRV and increase in HR, suggesting that even with the alcohol removed from the equation my sleep quality is worse during that night. This might be explained by the fact that this is also the night I’m most likely to spend with my girlfriend.
I don’t want to say that spending time with her is causing me to be more stressed (at least not most of the time), but I do sleep considerably better when I’m by myself.
Sleep scientist Matthew Walker, author of the excellent book Why We Sleep, confirms this and even suggests a “sleep divorce.” He and his long-term partner are practicing it themselves, sleeping in separate bedrooms and during their own ideal sleep window:
“It’s worked out wonderfully; we both feel a lot more rested, and our relationship is better in every way. Your desire to be intimate increases the more you sleep. Getting better and longer quality sleep raises testosterone levels, vital for both men’s and women’s sex drives. I believe we must spread this message and fight the stigma surrounding sleeping in separate beds – for the good of our collective health as well as our relationships.”
I know how miserable and awful I am to be around when I didn’t sleep well, so I fully support his idea of a sleep divorce, for everyone’s benefit.
The time I go to bed also varies somewhat predictably throughout the week.
Bedtime hour across weekdays. Left for all days, right only for days with no or low alcohol. Zero corresponds to midnight.
When including the nights I got out and drink, my weekends have by far the most variance and are shifted towards later sleep times.
Removing the days of moderate to high alcohol consumption, my average bedtime is much more stable throughout the week, sometime between 11 p.m. and midnight. In this case, the night from Saturday to Sunday actually has one of my earliest bedtimes (possibly also because I might have been out late and drinking the previous night, and am now having a recovery night).
Over the entire timeframe, my sleep has definitely moved to an earlier time.
Moving average of my bedtime, with a window size of 14 days.
Especially in 2020 I rarely went to bed past 1 a.m. and was usually in bed by midnight. Part of this is probably due to the fact that going out just wasn’t a thing in 2020, and in previous years the occasional night out might have pushed the moving average to later times.
While at a first glance this might look like a positive development, I’m not so sure about it. Maybe this positive association just comes from a culture that is built for and around morning people (which I am definitely not).
In June 2020 I started a more “traditional” job, where I often have meetings from 10 a.m. and as a result, need to get up by 8:30 a.m. if I still want to enjoy my morning routine of coffee, reading, and meditation before starting work. This in turn also moved my bedtime to an earlier hour.
While this might not seem like an issue or particularly early to most people (if anything, probably the opposite), I am really not a morning person and at least subjectively feel most rested when I sleep from around 1 am to 10 am. This could also be a contributor to my increased stress markers in 2020.
However, at least looking at each day as an individual data point, bedtime hour does, quite surprisingly, not seem to have any impact on my HRV.
Impact of bedtime on heart rate variability during sleep. There appears to be essentially no correlation between the two variables.
My total amount of sleep takes a somewhat similar pattern to the bedtime hour.
Total sleep duration across weekdays. Left for all days, right only for days with no or low alcohol. Zero corresponds to midnight.
If boozy nights are included, my weekend nights show significant variation. However, once they are removed that variation drops, and Saturday to Sunday night becomes the night of most sleep, with an average of around 9.5 hours.
In general, my sleep duration is very healthy at around 8 to 9 hours.
But I know just how important sleep is for me, and how poorly I perform on little sleep. Less than 8 hours and the next day won’t be great, less than 7 hours and it will be pretty miserable.
Looking at the moving average shows a similar picture.
Moving average of my total sleep duration, with a window size of 14 days.
It mostly stays above 8 hours, with 2020 actually appearing to be my year of longest overall sleep (which is again similar to Oura’s findings on their average user).
But all sleep is not created equal.
To understand a bit better at what’s actually going on, let’s take a look at the science of sleep phases.
Phases of Sleep
To really understand and appreciate what sleep is and how it supports our health and creativity, we have to look at our brains.
Quoting from my book Time Off, which has an entire chapter dedicated to sleep:
“Sleep is categorized into several different phases. A healthy person tends to go through several sleep cycles per night, each of which spans all the different phases and lasts for around 90 minutes. Simplifying things a little bit, the two primary phases of interest are deep sleep and rapid eye movement (REM) sleep. The former is largely responsible for physical healing and recovery, while the latter takes care of emotional healing and creativity.”
In addition, deep sleep is also responsible for memory consolidation. It’s the diligent librarian that makes sure nothing gets lost, everything is filed in the right place, and memories get efficient tags and cross-references.
REM sleep on the other hand is the phase during which dreaming occurs, and it’s a creative powerhouse, as well as a tool for us to process emotions. It makes us more emotionally stable, and as a result, more empathetic and understanding of others.
Everything that doesn’t fall into these two phases can, oversimplifying things a bit, be categorized as light sleep.
While few people question the accuracy of Oura’s heart rate and HRV measurements, there is some ongoing debate about how good it is at identifying the different sleep stages. However, even if the data is potentially not 100% accurate, as long as it is consistent there is still valuable.
So let’s take a look at how my ratios of different sleep phases change with the total sleep duration.
The ratio of different sleep phases in relation to total sleep duration.
While the amount of light sleep seems to be fairly independent of sleep duration, deep sleep and REM sleep definitely show some correlation.
The longer I sleep, the less deep sleep and the more REM sleep I seem to be experiencing.
This pattern is in good agreement with the science.
Our first sleep cycle is usually dominated by deep sleep. Then, as the night progresses, our sleep shifts more and more towards REM sleep. The final cycles of sleep, especially if we get a full night of sleep, are dominated by REM sleep.
Thus, many of us who cut our sleep short on a regular basis, consistently do not get the full benefits of REM sleep, including boosted creativity, better emotional regulation, and stress relief.
Another factor that can dramatically reduce the effectiveness of sleep, even if we get enough hours, is alcohol.
Impact of alcohol consumption on my sleep phases.
Alcohol clearly kills my deep sleep, leading to more light sleep instead.
Surprisingly, REM sleep seems to be more or less unaffected by alcohol. This goes against the academic literature, which has shown that alcohol has a particularly devastating effect on REM sleep.
Subjectively, I do remember more dreams on nights when I was drinking. However, my guess is that this is just due to lighter sleep and more frequent wake-ups, not due to actually dreaming more.
Unless I’m a strange medical outlier (unlikely), or the scientific literature is pretty wrong (even less likely), the data above might show some of the limitations of Oura’s sleep phase detection. It’s possible that some of the effects caused by alcohol look similar to Oura’s sensors as REM sleep does.
Not surprisingly, and as we have already indirectly seen above, drinking alcohol tends to shift my bedtime hour to later times (or said less causally, I tend to drink more on days on which I also go to bed late).
Relation of alcohol consumption and bedtime.
What is a bit surprising though is that my bedtime essentially seems to have no effect on my sleep phase ratios (plot not shown). From the literature, I would have expected earlier sleep times leading to more deep sleep, and later sleep times to more REM, but this was not the case.
What does show some impact on sleep ratios though is caffeine.
Impact of caffeine consumption on my sleep phases.
Similar to alcohol, more caffeine seems to increase light sleep at the expense of deep sleep, leaving REM mostly unaffected.
Just as for the heart metrics, the negative effects on sleep phases seem to come most strongly above a certain threshold, but even at lower values, there might be some effect here.
Surprisingly, sleep latency, the amount of time it takes me to fall asleep, seems to be almost unaffected by caffeine, even though that’s probably the first thing that would come to mind when thinking of an over-caffeinated day.
Impact of caffeine consumption on sleep latency.
However, as I’ve mentioned in my last article on this, sleep latency is a metric I’m particularly skeptical about and don’t think Oura is very good at detecting accurately. Especially long latencies seem to be drastically underestimated by the ring, probably due to me moving around in bed so much that it thinks I’m not even in bed yet.
Looking at each day in isolation shows some interesting correlations, but it also misses out on the time-series aspect of the data.
The moving averages can again give some insights into this.
Moving averages of my sleep phases, with a window size of 14 days.
Most obvious is that my deep sleep seems to have increased over 2019, in exchange for less light sleep, with a slight reversal in 2020.
What’s less obvious in the above plot, but somewhat more concerning, is that my REM sleep has been getting less and less over the years. The effect becomes clearer when plotting the total amount of REM sleep I’ve been getting each night by itself.
Moving averages of the amount of REM sleep I get every night, with a window size of 14 days.
I’ve noticed this for a while just by looking at my daily values and reflecting on how I feel, but seeing it plotted as a trend like this really drives home this point.
Whether or not Oura’s estimate of REM sleep is actually correct or not, I have subjectively noticed a strong correlation between my energy levels and emotional wellbeing and the amount of REM sleep Oura reports. So seeing this consistently decrease over time is a bit concerning, and might be partially responsible, or at least related to my higher perceived levels of stress and anxiety in 2020.
This is probably true for many people, but 2020 has been the first time in my 30 years so far that I felt genuinely anxious for extended periods of time.
Objectively, I’ve not been affected negatively by the pandemic. I never had to worry about job security or had to directly face any of the other stressors many people were confronted with. I’m sure there is some collective level of anxiety that didn’t go completely past me, but I don’t think all the changes I personally experienced can be blamed on the pandemic.
So this decrease in REM sleep (or whatever it is Oura is detecting as REM sleep), whether it’s cause or effect, is definitely something I’d like to keep an eye on, especially knowing how critical REM sleep is to emotional wellbeing.
Overall Conclusions
While I’ve presented a lot of data and analysis here, this is only the tip of the iceberg.
The Oura ring provides a lot of additional variables that I didn’t look into here, and the analysis itself was also fairly simplistic. But this simplicity is the best way to start gaining some valuable insights.
Overall, I want to be cautious reading too much into these results.
The relations of the different variables are so complex that it’s almost impossible to clearly identify any causality. And even the correlations that appear in the data might be driven by some hidden third variable, as we have seen when removing high alcohol days from the data.
Also, both my own self-tracking and modeling, as well as Oura’s measurements and models contain some uncertainty.
If there is one conclusion I’m confident to draw with very little doubt, it’s that – just as I’ve seen in my previous study with less data – alcohol beyond a small quantity is really bad for your sleep and recovery.
But despite the above caveats, I think even beyond this rather obvious observation there are some interesting conclusions to draw from the analysis.
For caffeine, the effect of my love for coffee is not as bad as I had feared. Above around 70 mg of caffeine in my blood at bedtime, I risk higher resting HR, lower HRV, and more light sleep. But as long as I keep my caffeine levels below this fairly high threshold, I shouldn’t have to worry too much about affecting my sleep in a negative way.
I’m again wondering if this raises the question if a lot of medical studies have even looked at this lower (but more realistic) range, or just extrapolated that caffeine is generally bad for sleep from high caffeine doses administered close to sleep. Maybe someone can point out that I just didn’t look at the right studies.
The results of the high caffeine days raise another question pointing at the complexity of these variables: was the high caffeine actually the cause of my light sleep and lower HRV, or were those days particularly stressful days which made me reach for more caffeine, and it was actually the stress that caused the bad results more so than the caffeine itself?
The same could be said for alcohol. Maybe days on which I was already stressed made it more likely for me to drink? Again, we see how complicated and entangled these variables are.
Attempting to answer this particular question would require to also manually track my perceived stress levels throughout the days. Maybe an experiment for the future.
Overall 2020, or at least the part for which I had data, didn’t look too good for me, with many stress indicators seeming to have gotten worse over the year. As I mentioned, I’m aware of some very concrete stressors, and working on them will be a big focus for me in 2021.
I’m curious to see how this will be reflected in my data going forward from here.
Concretely, I want to increase my amount of REM sleep again, and also work on my heart metrics, decreasing resting heart rate, and increasing HRV.
I also want to try to flatten my heart rate and HRV throughout the week. I didn’t expect stress to increase so noticeably throughout the week, and I don’t think it has to be that way.
I have some experiments in mind, and would also love to hear your suggestions for hitting my goals.
Happy self-tracking and bio-hacking!