Last year, I posted the results of a self-experiment on Melatonin. This year, I ran another self-experiment from January to June, the results of which I'll lay out in this post.

Tl;dr: I randomized three interventions on a daily basis and tracked several metrics throughout the day to see what the short-term effects are. Interventions centered on coffee intake, meditation, and doing jumping jacks, all of which happened primarily in the morning. Caveats apply (n = 1, nothing was blinded, some analyses slightly differ from my "preregistration"). The main results:

• Caffeine had a range of positive effects on me:
  • Improved mood (+0.8 on a 0-10 scale, p < 0.0001) 
  • Improved energy level (+1.0 on a 0-10 scale, p < 0.0001)
  • Reduced distraction (-0.7 on a 0-5 scale, p < 0.0001)
  • Across all days, productivity rose by +9% through caffeine, but insignificant (p = 0.38); I then performed a sensible-but-not-preregistered little p-hack, filtering for the types of days where the my productivity metric is most meaningful, and obtained +19% (p = 0.01)
  • Improved subjective job performance (+0.5 on 0-5 scale, p = 0.009)
• Caffeine intake slightly delayed time to fall asleep (from 28 to 31 minutes on average, p = 0.014) and perhaps reduced subjective sleep quality (7.3 → 7.1 on a 0-10 scale, p = 0.21)
• Decaf showed ~none of the above effects, with the possible exception of productivity, which saw an increase of +20% across all days compared to the control group, but not statistically significant (p = 0.18)
• Doing 50 jumping jacks in the morning may perhaps improve my productivity (+20%, p = 0.03) (but this turns to +14%, p = 0.08 when applying the sensible-but-not-preregistered adjustment mentioned above)
• 10-20 minutes of meditation a day had no measurable short-term effects on anything

I performed the analysis by hand via hypothesize.io. Zero AI involvement went into the analysis or this post. (Although I'm not even sure if that's necessarily a good thing at this point)

Experiment Setup

I ran an experiment for the first six months of 2026 where I randomized three different interventions independently of one another on a daily basis. At the beginning of each day, I would (before looking up that day's randomized group assignment) decide which of the three interventions I'd perform, with the goal of doing all three every day, which sometimes wasn't possible, e.g. due to early appointments, traveling, or being sick. On 136 days, I followed through for at least one of the three interventions. On 133 days, I did all three.

The three experiments were:

1. Caffeine, randomized between four equally likely groups (done on 135 days)
   1. Control - no caffeine, no decaf (38 days)
   2. Decaf - around 200ml of decaf coffee (23)
   3. Coffee - around 100ml of regular coffee or ~120mg of caffeine (presumably) (36)
   4. Coffee - around 200ml of regular coffee or ~240mg of caffeine (38)
   5. Coffee and decaf were consumed with oat milk and no added sugar.
2. Jumping jacks (135 days)
   1. Control - none (63)
   2. 50 jumping jacks in the morning (72)
3. Meditation (134 days)
   1. Control - none (45)
   2. 10 minutes in the morning (46)
   3. 10 minutes in the morning plus 10 minutes in the afternoon (43)

Independent of the coffee experiment, I also had one bag of green tea each morning (~50mg caffeine, presumably), so my caffeine control days were technically not free of caffeine.

Meditations were performed in the form of different guided meditations, all of which were about 10 minutes in length.

Why three separate experiments? I came to the conclusion that running these in parallel would yield higher power than running them sequentially, and also just found it more interesting this way. 

None of the interventions were blinded. For decaf vs. regular coffee, it would have been possible in principle, but the effort of coming up with a good blinding strategy seemed high and was likely a fake alternative that would just have caused me to postpone the experiment endlessly. But this does, of course, add some grains of salt to the coffee findings, particularly given that many of my metrics are subjective self-reports.

What I Measured

I measured the following metrics on a daily basis:

• Which of the three interventions I had performed according to the randomization
• Sleep quality (0-10) and how long I took to fall asleep (in minutes), both being subjective estimates
• On work days: subjective estimate of my performance on the job (0-5)
• A metric for my productivity during my non-job time. This one is relatively objective and standardized (I've been tracking this for about ten years now) and coupled to actual things I do throughout the day. Yields an unbounded non-negative number that ranged from 0 to 41 with a mean of 16.1.
• Subjective energy level at 11am and 3pm (0-10)
• My stress level (0-5)
• My general mood (0-10)
• My general level of distraction (0-5)

The scales that I used for many of the subjective estimates are somewhat arbitrary, some from 0-5 and some from 0-10, and are based on what I'm used to from earlier experiments and established tracking habits.

Expectations

I shared a google doc with my plans and expectations with a few friends before I started the experiment — not quite a preregistration, but close enough, I guess. These were the effects I specifically wanted to look for (not necessarily expecting them to exist, let alone end up as statistically significant, but that I explicitly wanted to check once I had the data):

• Coffee
  • Improves mood
  • Improves personal productivity
  • Improves job performance
  • Improves energy level
  • Reduces distraction
  • Increases time to fall asleep
  • Decaf does not improve my productivity
  • Decaf does not improve my mood
  • Decaf does not improve my energy level
  • More coffee (2 cups) improves productivity more than less (1 cup)
• Meditation
  • Improves mood
  • Improves productivity
  • Improves job performance
  • Reduces distraction
  • Reduces time to fall asleep
• Jumping jacks
  • Improves mood
  • Improves productivity
  • Improves energy level

Consider yourself nudged to write down some predictions before reading on, e.g., what types of effect sizes you would expect for any of these.

Results

Going through the list of expected effects, here are the effect sizes I found, with significance levels (not adjusting for multiple comparisons). For all of these, I just evaluated the given experiment in isolation.

All p-values are based on two-sided t tests.

• Coffee
  • Improves mood (134 days^[1]):
    • Caffeine intake (experiment groups 2 and 3) vs. none (groups 0 and 1) is associated with a mood increase from 5.6/10 to 6.4/10 (+0.8, p < 0.0001)
    • Looking at all 4 groups, I get:
      • Control: 5.7 (on 0-10 scale)
      • Decaf: 5.3 (-0.4, p = 0.13)
      • 1 Cup: 6.2 (+0.5, p < 0.01)
      • 2 Cups: 6.7 (+1.0, p < 0.001)
    • However, keep in mind that coffee intake was not blinded, and mood is a subjective metric.
    • Confirmed ✅
  • Improves personal productivity:
    • Taking all 135 days into account, caffeine intake is associated with an increase in average productivity from 16 to 17.5 (+9%, p = 0.38)
    • If I filter for off days (63), which I did not preregister but which in hindsight does make some sense, as they carry way more signal here (because this productivity metric only measures how I spend my non-working time, and working days leave little time and energy to be productive in other ways), I get my average productivity increasing from 21 to 25 (+19%, p = 0.01)
    • Directionally confirmed, and specifically confirmed for non-working days (✅)
  • Improves job performance (66 working days):
    • Caffeine intake is associated with an increase in subjective rating of job performance on a day from 2.8 to 3.3 (+0.5 on 0-5 scale, p = 0.009)
    • Confirmed ✅
  • Improves energy level (134 days)
    • Energy at 11am: 6.6 (with caffeine) vs 5.6 (no caffeine), so +1.0 on a 0-10 scale (p < 0.0001)
    • Energy at 3pm: 6.7 vs 5.7, so +1.0 on a 0-10 scale (p < 0.0001)
    • Confirmed ✅
  • Reduces distraction (134)
    • Distraction on caffeine days averages 2.1 compared to 2.8 on non-caffeine days, -0.7 on a 0-5 scale (p < 0.0001)
    • Confirmed ✅
  • Increases time to fall asleep (126 days for which I have the data; I guess this has a chance of introducing bias: sometimes I failed to track my sleep data the next day, which may or may not be influenced by my sleep quality that night and thereby by caffeine intake the day before)
    • Time to fall asleep had two outliers (50 and 70 minutes) that were more than 5 standard deviations above the median. I accepted hypothesize.io's suggestion to clip these outliers to 5 standard deviations (9 to 49 minutes).
    • Caffeine is associated with an increased time to fall asleep from 28 to 31 minutes (+11%, p = 0.014)
    • Unfortunately, I did not track what time I went to bed or how long I slept, which might be relevant details
    • Confirmed ✅
  • Impacts on sleep quality
    • This was not part of my preregistration; not sure why not, but I now tested it anyway (sue me)
    • Caffeine intake the previous day is associated with a subjective sleep quality decrease from 7.3 to 7.1 (-3%, p = 0.21)
    • Notably, I usually limited my coffee intake to between 8am and 12pm to reduce negative impacts on sleep quality
    • (Not pre-registered, expected direction but not significant ❓)
  • Decaf does not improve my productivity
    • I looked at the 61 days where I either had decaf or the control group (i.e., excluded all caffeine days)
    • Taking all these days into account, decaf was associated with an average productivity of 18 vs 15 in control (+20%, p = 0.18)
    • Limiting this to the 32 non-working days, I find an increase from 20.5 to 22.5 (+10%, p = 0.45)
    • Definitely an interesting data point; perhaps the non-caffeine-related positive effects of coffee do have some impact on (my) productivity? But then again, with the limited number of days, we end up with too high p-values to take this seriously. More research is needed, I guess, but I may not be willing to randomize coffee vs. decaf for long enough to find out more.
    • Unclear ❓
  • Decaf does not improve my mood (61 days)
    • Decaf mood was 5.3 compared to 5.6 in control (-5%, p = 0.15)
    • Confirmed ✅
  • Decaf does not improve my energy level
    • 11am: 5.6 (control) -> 5.4 (-4%, p = 0.36)
    • 3pm: 5.75 (control) -> 5.6 (-3%, p = 0.41)
    • Confirmed ✅
  • More coffee (2 cups) improves productivity more than less (1 cup)
    • Looking at all 74 days on which I took either 1 cup or 2 cups of coffee
    • Average productivity for 1 cup was 16.5, vs 19 for 2 cups (+15%, p = 0.23)
    • Limiting this to non-working days (31), I'm left with average productivity of 24.5 (1 cup) and 26 (2 cups) (+6%, p = 0.53)
    • Expected direction but not significant ❓
• Meditation
  • Note: I had 3 groups for meditation - control, a single session in the morning, and two sessions throughout the day; in practice, I often failed to do the second session, but reliably did the first one. Hence, for the analyses here, I just differentiate between control and "any meditation". The data is not extensive enough for a meaningful dose-response test either.
  • Improves mood (132 days)
    • 5.9 (Control) vs 6.05 (+0.15 on 0-10 scale, p = 0.43)
    • Not confirmed ❌
  • Improves productivity
    • 17 (control) vs. 17.5 (+3%, p = 0.673)
    • On non-working days (62), it's 22.5 (control) vs 25 (+11%, p = 0.18)
    • Directionally confirmed but not significant ❓
  • Improves job performance (65 days)
    • 3.1 (Control) vs 3.09 (-0.01 on a 0-5 scale, p = 0.97)
    • Not confirmed ❌
  • Reduces distraction
    • 2.5 vs 2.5 (p = 1)
    • Not confirmed ❌
  • Reduces stress
    • This was not part of my preregistration; not sure why not, but I now tested it anyway
    • 0.75 (Control) vs 0.76 (+0.01 on a 0-5 scale, p = 0.92)
    • Not confirmed ❌
  • Reduces time to fall asleep
    • Time to fall asleep had two outliers (50 and 70 minutes) that were more than 5 standard deviations above the median. I accepted hypothesize.io's suggestion to clip these outliers to 5 standard deviations (9 to 49 minutes).
    • 28.5 (Control) vs. 29.5 (+1min, p = 0.25)
    • Not confirmed ❌
• Jumping jacks
  • Improves mood (135 days)
    • 5.9 (Control) vs 6.0 (+0.1 on 0-10 scale, p = 0.59)
    • Not confirmed ❌
  • Improves productivity
    • Over all 135 days, doing jumping jacks in the morning is associated with an increase in productivity from 15 to 18 (+20%, p = 0.03)
    • On non-working days only (63), we get 21.5 (Control) -> 24.5 (+14%, p = 0.08)
    • Surprisingly, kind of confirmed ✅
    • (but let's be honest, the chances of 50 jumping jacks in the morning leading to +20% productivity seem rather slim; I'd expect this to be a false positive, at least in this magnitude — maybe some n=2nd person wants to run a follow-up?)
  • Improves energy level
    • 11am: 6.0 -> 6.0
    • 3pm: 6.3 (Control) -> 6.2 (-0.1 on 0-10 scale, p = 0.66)
    • Not confirmed ❌

Big Regression Model

Finally, I also ran a few bigger regression models with many of the factors taken into account at once.

I limited this to the days where I had performed all three experiments (133 days).

My independent variables were:

• Caffeine taken (1 or 2 cups) or not (decaf or control)
• Meditated (1 or 2 sessions) or not
• Jumping jacks done or not
• A measure of caffeine intake of the past days (taking the cups of actual coffee of the last three days, weighted by 0.5 for yesterday, 0.25 the day before, and 0.125 the day before that, yielding a range from 0 to 1.75; median was 0.5625, mean 0.6813) to quantify possible short-term tolerance effects

Below, one subsection follows for each dependent variable I looked into.

Productivity

For this analysis, I had one additional independent variable to the ones mentioned above:

• Work day vs off day (on a numeric scale of 0-1, where I had some half work days logged as 0.5)

The reason being that my productivity tracking refers to how I'm spending my free time, so naturally working days have much lower values, typically, than off days.

Here, the effects of caffeine on productivity are significant at +2.8, p = 0.01.

It's not surprising that work days have a huge negative effect; I'm rather surprised the effect isn't larger.

Jumping Jacks (second bar) look promising again, but are not significant with p = 0.09.

Interestingly, the amount of caffeine taken on past days has zero effect here. Subjectively, I often feel that after two days of having coffee, I end up having a rather challenging day. I would have expected to see this reflected in the productivity data.

Work Performance

This analysis only entailed 65 days, so naturally p-values are larger. The y-axis is relative to a subjective 0-5 scale.

Caffeine still wins (p = 0.018). Short-term caffeine tolerance effects are basically 0. Meditation is close to 0, as in so many of my analyses (p = 0.63). The previously observed positive effect of jumping jacks entirely disappears (p = 0.88).

Mood

Based on 131 days.

Caffeine takes the lead once again, with p < 0.0001.

Finally, we also see something similar to a tolerance effect, with the weighted caffeine amount from the previous days having a slight negative effect of around -0.23 on a 0-10 scale (p = 0.003).

Here, meditation and jumping jacks turn out to be leaning positive but non-significant (p = 0.23 for meditation and p = 0.19 for jumping jacks).

Work days show a tiny negative effect (p = 0.08).

Meta Learnings

Was this experiment a good use of my time? Did I learn anything I didn't already know? LessWrong user siarshai argued that self-experiments are often not very useful, and I partially agree. E.g., I'm not at all surprised about coffee showing all these positive effects. I already knew it improved my mood and was very certain it increased my productivity. Still, I do take away a few updates from this:

• Jumping jacks may be a good idea. I doubt my effect size would hold up in a longer trial, but I guess my results here do make it more likely that I'll actually establish this as a habit to overcome sleep inertia in the morning, as it only takes a minute and doesn't seem to have any downsides.
• I'm surprised meditation showed basically nothing in most of these analyses. Not finding sufficiently large effects to achieve statistical significance was mostly priced in, but I didn't even really find non-significant positive effects. My working model of how meditation affects me was basically that I assumed it reduced my stress level and brought me into a less impulsive and more reflective state of mind, which, hypothetically, should have had some positive effects on the rest of the day. Of course, meditation may still have longer-term benefits that I can't possibly measure in an experiment that randomizes on a daily basis. Still, this lets me update away slightly from maintaining a daily meditation habit.
• My finding that decaf was associated with +20% productivity (p = 0.18) seems somewhat interesting. I'd assume that it's a fluke and would regress towards 0 in a longer experiment. But it's conceivable that the positive effects on productivity are partially not due to the caffeine, after all. It could be anything from the "ritual" to the positive non-caffeine-related effects of coffee to better hydration. So, maybe this is reason enough to run a follow-up, although I'm not so sure it's worth it.
• Interestingly, I find that running these self-experiments also helps me build (or at least try) new habits. If I just decided to do jumping jacks in the morning, chances are I wouldn't stick to that, at least not without a fight. But if I plan to run a months-long self-experiment, then this experiment carries some authority, and I'm more likely to do what it tells me to do than I might otherwise.
• I think next time I'm about to analyze an experiment, I'll prepare the entire analysis on shuffled data (basically the experiment results, but with dates/experimentations shuffled, so that all effect sizes would be just noise around 0). This would allow me to identify all the structural challenges of the analysis without seeing any concrete outcomes yet. Once I'm happy with the analysis, I can then plug in the real data and just rerun the code. This seems better than what I tried here, namely (privately) preregistering my analysis in somewhat vague language, and then, while analyzing the data, occasionally realizing that some degrees of freedom were left.

1. ^{^}

   The day counts here slightly differ from those reported earlier in Experiment Setup, as there turned out to be two days where I technically did the experiments, but forgot to track some metrics later in the day