**The most counterintuitive finding is that late caffeine, the habit people blame most, has almost no link to how long people sleep. It is tied to how well they sleep.** Across all 8,500 users, caffeine after 5pm (`caffeine_post_5pm_mg`) has these correlations (step 1 and 2 results): - **Total sleep hours:** r ≈ **-0.04**, essentially zero. - **Sleep latency:** r ≈ **+0.20**, a modest increase in time to fall asleep. - **Deep sleep %:** r ≈ **-0.34**, a clear reduction. This is the second-strongest negative link to deep sleep that I computed. So people who drink more caffeine late in the day sleep about as long, but take longer to fall asleep and get less deep sleep. A "caffeine cutoff" rule judged on sleep duration alone would find nothing. That is easy to miss, and it contradicts the common assumption that caffeine mainly costs you hours of sleep. Other results that cut against expectations: - **Physical activity** is unrelated to sleep and fatigue: r ≈ 0.004 with sleep hours and -0.03 with fatigue. - **Blue light filters** show little relationship with latency (r ≈ -0.06) or sleep hours (r ≈ 0.04). Filter users have only slightly better averages (step 3 result): | | Filter on | Filter off | |---|---|---| | Sleep latency (avg min) | 39.6 | 41.6 | | Sleep hours (avg) | 6.33 | 6.21 | | Next-day fatigue (avg) | 3.58 | 3.99 | - **Screen brightness** is only weakly tied to latency (r ≈ 0.09), while **bedtime phone minutes** are tied very strongly (r ≈ 0.86). Time on the phone appears to matter far more than how bright or filtered the screen is. These are correlations from observational data, so they show association, not cause. I only computed the Pearson correlations and filter-group averages above, so I did not test whether the caffeine effect holds within subgroups.