**The most counterintuitive finding is that the usual "sleep hygiene" fixes barely register in this data, while minutes of phone use at bedtime dominate everything.** The blue-light filter is the clearest example. **What is strong (the expected part, but extreme):** - Bedtime phone minutes correlate **+0.86** with sleep latency and **+0.71** with next-day fatigue. - They correlate **−0.55** with total sleep hours. - Sleep hours correlate **−0.88** with fatigue, and morning alarm snoozes correlate **+0.88** with fatigue. **What is surprising (the near-zero effects):** - **Blue-light filters barely help.** 3,976 of the 8,500 users (47%) have a filter on. Their correlation with latency is only −0.057 and with sleep hours only +0.039. - Filter users average 39.6 min of latency against 41.6 min for non-users, about 2 minutes. - They sleep 6.33 h against 6.21 h, about 7 minutes more. - Their average fatigue is 3.58 against 3.99, which is a modest gap. - **Screen brightness is nearly irrelevant** to latency (r = +0.09). - **Physical activity shows essentially no link** to sleep hours (r = +0.004) or next-day fatigue (r = −0.03). - **Late caffeine does not shorten sleep** (r = −0.038 with sleep hours), yet it clearly hurts sleep quality. It correlates **−0.34** with deep sleep, +0.20 with latency and +0.18 with fatigue. Caffeine seems to change how well people sleep, not how long. **How to read this:** the data points to the amount of time spent on the phone, not how the screen is configured, as the driver of sleep debt. A filter-based recommendation would have little support here. Advice to cut bedtime phone minutes and caffeine would have far more. These are correlations from observational data, so they don't prove cause. The filter's weak link could partly reflect who chooses to use one. The correlations were computed with SQL over `raw.bedtime_screentime_sleep_debt`, and the filter comparison comes from the grouped averages by `blue_light_filter_active`.