Three questions stand out. The correlations below come from queries I ran across all 8,500 users. They are observational, so they show association, not cause. **1. How much of the sleep-debt and fatigue problem runs through bedtime phone use?** - Bedtime phone minutes and sleep latency are very strongly related (r ≈ 0.86). - More phone time also goes with fewer total sleep hours (r ≈ −0.55). - Latency tracks next-day fatigue (r ≈ 0.74), and total sleep hours track it even more tightly (r ≈ −0.88). - A plausible chain is phone time → longer time to fall asleep → less sleep → more fatigue. - This is the most actionable question. The follow-up is to test whether the chain holds within segments (app, occupation, chronotype) or whether it is a single population-wide effect. The pre-aggregated gold tables by app, occupation and chronotype can answer that directly. **2. Do blue-light filters and lower brightness actually help, or is that a myth?** - Users with a filter on average 39.6 minutes of sleep latency, against 41.6 minutes with it off. That is about a 2-minute gap, and the correlation is weak (r ≈ −0.06). - Screen brightness is also only weakly related to latency (r ≈ +0.09). - Phone minutes dominate latency (r ≈ 0.86), so duration looks far more important than screen settings. - The open question is whether filters help specific groups, such as evening chronotypes or particular apps. It is also worth checking whether filter users simply spend less time on their phones, which would confound the 2-minute gap. **3. Does late caffeine hurt how well people sleep rather than how long?** - Caffeine after 5pm is almost unrelated to total sleep hours (r ≈ −0.04). - It is moderately negatively related to deep sleep percentage (r ≈ −0.34) and positively related to sleep latency (r ≈ 0.20). - A caffeine cutoff may therefore matter for sleep quality, even though it would not show up in sleep-debt categories, which are derived from sleep hours alone. - For comparison, physical activity has essentially no relationship with latency (r ≈ −0.01). - The follow-up is to look for a threshold dose and to check whether the effect differs by chronotype or occupation. All three questions can be answered from this table. The caveat is that sleep_debt_category is derived directly from total_sleep_hours, so it should not be treated as an independent outcome when modeling.