**The most counterintuitive finding is that how long people use their phone at bedtime matters far more than the light the screen gives off. The usual remedies, a blue-light filter, lower brightness, no late caffeine and exercise, barely register against sleep hours.** - **Phone time dominates.** Bedtime phone minutes correlate at r = 0.86 with sleep latency, 0.71 with next-day fatigue and -0.55 with total sleep hours (n = 8,500). A correlation of 0.86 on behavioural data is unusually strong. - **Blue-light filters do almost nothing.** - The correlation with latency is -0.057 and with sleep hours +0.039. - Filter users (3,976) fall asleep in 39.6 minutes on average, versus 41.6 for non-users (4,524). They sleep 6.33 hours versus 6.21. - That is a gap of about 2 minutes and 7 minutes of sleep. Fatigue is somewhat lower for filter users (3.58 vs 3.99). - Even if filters help, they do not offset the effect of phone time. - **Screen brightness is nearly irrelevant.** Its correlation with latency is only 0.09. - **Caffeine hits sleep quality, not duration.** - After-5pm caffeine is essentially uncorrelated with total sleep hours (-0.04). - It is, however, clearly linked to less deep sleep (-0.34), longer latency (+0.20) and higher fatigue (+0.18). - A "caffeine cutoff" would therefore show up in sleep quality rather than in hours slept. - **Exercise shows nothing.** Physical activity correlates about 0 with sleep hours (0.004) and fatigue (-0.03). Two caveats: - The data is observational, so these are associations, not proven effects. - Fatigue tracks sleep hours (r = -0.88) and alarm snoozes (r = 0.88) very tightly. Together with the near-perfect phone-time signal, this suggests the relationships are cleaner than real-world data usually is, so I would treat the magnitudes with some caution. Everything above comes from the correlation queries (steps 0–12) and the filter-versus-no-filter comparison (step 3).