## Most surprising finding: blue-light filters barely matter, even though phone use matters a lot This is a partial answer, because the step limit was reached before I could test some follow-ups. The evidence still points clearly in one direction. ### What the data shows (n = 8,500) | Relationship | Correlation (r) | |---|---| | Phone use vs. sleep latency | **+0.858** | | Phone use vs. next-day fatigue | **+0.711** | | Phone use vs. total sleep hours | −0.554 | | Blue-light filter vs. sleep latency | **−0.057** | | Blue-light filter vs. sleep hours | +0.039 | | Blue-light filter vs. fatigue | −0.077 | | Screen brightness vs. latency | +0.093 | Phone use is strongly tied to taking longer to fall asleep (r ≈ 0.86). That is unusually high for behavioral data. The blue-light filter, the obvious fix, shows almost no relationship with any sleep outcome. The group averages tell the same story. Filter users (3,976 rows) and non-users (4,524 rows) differ only slightly: | Measure | Filter on | Filter off | |---|---|---| | Average sleep latency | 39.6 min | 41.6 min | | Average sleep hours | 6.33 | 6.21 | | Average fatigue score | 3.58 | 3.99 | | Average deep sleep | 22.1% | 21.4% | These are small gaps, roughly 2 minutes of latency and 0.4 points on a 10-point fatigue scale. They are far smaller than the swings associated with phone use itself. ### Other near-surprises - **Caffeine after 5pm barely relates to total sleep** (r = −0.038). It does relate to less deep sleep (r = −0.343) and, more weakly, longer latency (+0.199) and higher fatigue (+0.181). Caffeine seems to change sleep quality more than sleep quantity. Median post-5pm caffeine is 0 mg, so most people have none, which may dampen the correlations. - **Physical activity shows essentially no relationship** with sleep hours (r = 0.004) or fatigue (r = −0.029). - **Snoozes track fatigue (r = +0.881) as tightly as sleep hours do (r = −0.879).** Snoozing is probably a symptom of tiredness, not a cause. ### Caveats - These are correlations only, so none of this shows causation. - I did not test whether filter users differ in phone use. A filter might help only heavy users, and a pooled correlation would hide that. - I did not run significance tests or control for age or other variables. With n = 8,500, even r ≈ −0.06 is probably distinguishable from zero. But it is tiny in practical terms, explaining well under 1% of the variance. - The filter variable is binary, so its correlations are less directly comparable to those of the continuous variables. **Bottom line:** the data suggests that how much people use their phone matters far more for sleep than whether a blue-light filter is on. That is a counterintuitive result for a widely used remedy.