Three questions stand out. The data supports the first two with computed numbers. The third is a promising direction I checked only in part. ## 1. How much of the fatigue problem runs through bedtime phone time? Bedtime phone minutes is the strongest signal in the data, and the chain it feeds is tight (all 8,500 users, Pearson r): | Link | r | |---|---| | Phone minutes → sleep latency | **0.86** | | Phone minutes → total sleep hours | −0.55 | | Phone minutes → next-day fatigue | 0.71 | | Sleep latency → next-day fatigue | 0.74 | | Total sleep hours → next-day fatigue | **−0.88** | The pattern is that more phone time goes with a longer time to fall asleep, less total sleep, and more fatigue. Sleep hours are the closest predictor of fatigue. This points to a mediation question: how much of the phone-to-fatigue link is explained by lost sleep rather than by phone use itself? Because the data is observational, this shows association, not proof of cause. ## 2. Do the usual "fixes", blue light filters and lower brightness, matter next to phone time? They barely register in these results: - **Blue light filter:** the correlation with latency is only −0.057. Average latency is about 39.6 minutes with the filter on and 41.6 without, a gap of roughly 2 minutes. - **Screen brightness:** the correlation with latency is only 0.093. Compared with the 0.86 for phone minutes, how long people use the phone looks far more important than how they use it. That matters for deciding what to recommend, since advice to switch on a filter may do little on its own. I did not control for phone minutes, so users with filters on could differ in other ways. ## 3. Does late caffeine hurt sleep quality rather than sleep length? Caffeine after 5pm is almost unrelated to total sleep hours (r = −0.038). It is more strongly related to deep sleep (r = −0.34) and shows a modest link to latency (r = 0.20). So a caffeine cutoff may be better justified by sleep quality than by duration. I'd test this properly with dose bands, and by checking whether the effect holds after accounting for phone time. ## A note on method `sleep_debt_category` is computed directly from `total_sleep_hours`. Using it as an independent outcome would partly repeat the sleep-hours result. A natural follow-up is to see which segments (bedtime app, occupation, chronotype) are most affected. The gold tables already break the data down by those groups, but I haven't queried them here.