Here are the three questions I'd pick. The first two rest on correlations I computed across all 8,500 users. The third rests on one caffeine result, and the segment comparison it points to has not been run. ## 1. How much of the sleep-debt-to-fatigue chain starts with bedtime phone minutes? Phone use at bedtime has the strongest link in the data to how long people take to fall asleep (r ≈ 0.86). It is also moderately linked to shorter total sleep (r ≈ -0.55) and to next-day fatigue (r ≈ 0.71). Fatigue then tracks the sleep outcomes themselves very closely: - Total sleep hours vs fatigue: r ≈ -0.88. - Sleep latency vs fatigue: r ≈ 0.74. This suggests a chain of phone minutes, then slower sleep onset, then fewer hours, then more fatigue. A mediation-style analysis could test how much of the phone-to-fatigue effect passes through latency and sleep hours. It could also test whether any threshold in phone minutes (the range is 1–180) is where sleep debt categories begin to worsen. ## 2. Do blue light filters and lower brightness actually help? This matters for product and advice decisions, and the data looks surprisingly unsupportive of the usual fix: - Users with the filter on average about 39.6 minutes of sleep latency, against about 41.6 minutes without it. That is a small gap of roughly 2 minutes. - The filter's correlation with latency is only r ≈ -0.06. - Screen brightness vs latency is r ≈ 0.09, also weak. Because phone minutes dominate (r ≈ 0.86), the useful comparison is filter effects within phone-use bands. That would show whether filters help at all once duration is held constant, or whether they only give a false sense of safety. Segment gold tables, such as the one by `primary_bedtime_app` and `blue_light_filter_active`, could show where filters matter most. ## 3. Does caffeine after 5pm hurt sleep quality rather than sleep duration? Caffeine after 5pm shows almost no linear relationship with total sleep hours (r ≈ -0.04). But in my first query, its correlation with deep sleep percentage was r ≈ -0.34, and its correlation with sleep latency was r ≈ 0.20. So the caffeine cutoff question may be about sleep architecture and onset rather than hours slept. This is worth checking with dose bands, for example whether a threshold in mg separates normal from impaired deep sleep. It would also be worth checking whether the effect differs by chronotype. ## Caveats - The data is observational, so none of these correlations prove causation. - Segment-level differences by app, occupation and chronotype are a natural follow-up, but I haven't tested them here. - `bedtime_phone_minutes` is flagged as person-identifying, so any analysis should stay aggregated, as the correlations above are.